You are on page 1of 118

METAPTUQIAKH DIATRIBH

ANAGNWRISH MOUSIKOU EIDOUS ME


TANUSTES

y V
z
x

Emmanou l Mpenèto AEM: 112


Epiblèpwn: Kwnstantno Kotrìpoulo

ARISTOTELEIO PANEPISTHMIO JESSALONIKHS

TMHMA PLHROFORIKHS - KATEUJUNSH YHFIAKWN MESWN

AKADHMAÛKO ETOS 2006-2007


Abstra t

The eld of automati musi al genre lassi ation on a multilinear e-


nvironment is examined in the present thesis. Most musi al genre las-
si ation te hniques employ ma hine learning algorithms for lassifying
ve torized data into genre labels. In the proposed approa h, ea h re or-
ding is represented by a feature matrix over time. Thus, a feature tensor
is reated by on atenating the various feature matri es. The eld of ten-
sor analysis is examined and a novel algorithm for tensor de omposition
is proposed. The te hnique is alled non-negative tensor fa torization
(NTF) and aims at de omposing an N -dimensional tensor into a sum of
elementary rank-1 tensors. Several variants of the NTF algorithm are de-
veloped, as well as a supervised lassi er is proposed. A variety of sound
des ription features are extra ted from 1000 re ordings overing 10 genre
lasses. The features in lude spe tral, temporal, per eptual, energy, and
pit h des riptors. In addition, several ma hine learning algorithms are
tested for omparative purposes. Experimental results rea h 75% gen-
re lassi ation a ura y and indi ate the superiority of the multilinear
algorithms over the linear ones.

ii
Mèlh trimeloÔ epitrop  : Kwnstantno Kotrìpoulo , Nikìlao Lˆskarh , kai Iwˆnnh

P ta .

To akadhmaðkì èto 2006-2007 o Mpenèto Emmanou l, me thn idiìthta tou w me-

taptuqiakoÔ foitht , dietèlese upìtrofo tou KoinwfeloÔ IdrÔmato Alèxandro S.

Wnˆsh .

iii
iv
Perieqìmena

1 Eisagwg  1
1.1 Stìqoi diatrib  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Dom  diatrib  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Logismì Tanust¸n 5
2.1 DianÔsmata kai Pnake . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1.1 Orismo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1.2 Idiìthte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Orismì Tanust¸n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.3 Prˆxei me Tanustè . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.3.1 Basikè Prˆxei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3.2 Ginìmena Tanust¸n . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3.3 Monadiao kai Isotropikì Tanust  . . . . . . . . . . . . . . . . . . 10

2.3.4 Sustol  Tanust  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3.5 Ginìmeno Tanust  me Pnaka . . . . . . . . . . . . . . . . . . . . . . . 12

2.3.6 Bajmwtì Ginìmeno . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3.7 Orjogwniìthta Tanust¸n . . . . . . . . . . . . . . . . . . . . . . . . 14

2.3.8 Anˆptugma Tanust  . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.3.9 Bajmì Tanust  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.3.10 Uper-summetriko Tanustè . . . . . . . . . . . . . . . . . . . . . . . 20

2.3.11 Tanustè w Grammiko Metasqhmatismo . . . . . . . . . . . . . . . . 21

2.4 Statistikˆ Mètra Uyhlìterh Tˆxh . . . . . . . . . . . . . . . . . . . . . . 22

2.4.1 Ropogenn trie kai SusswreÔtrie . . . . . . . . . . . . . . . . . . . 22

2.4.2 Idiìthte Rop¸n kai Susswreutri¸n . . . . . . . . . . . . . . . . . . . 24

2.4.3 Upologismì Statistik¸n Mètrwn Uyhlìterh Tˆxh . . . . . . . . . 26

2.5 Apoklsei Bregman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3 Anagn¸rish MousikoÔ Edou 31


3.1 Eisagwg  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

v
PERIEQŸ
OMENA PERIEQŸ
OMENA

3.2 SÔnola dedomènwn kai Ierarqe . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.3 Qarakthristikˆ Perigraf  MousikoÔ Edou . . . . . . . . . . . . . . . . . 36

3.3.1 Qarakthristikˆ qroiˆ . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.3.2 Rujmikˆ-qronikˆ qarakthristikˆ . . . . . . . . . . . . . . . . . . . . . 38

3.3.3 Armonikˆ kai melwdikˆ qarakthristikˆ . . . . . . . . . . . . . . . . . 39

3.4 Algìrijmoi Anagn¸rish . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.4.1 Mèjodoi qwr epbleyh . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.4.2 Mèjodoi me epbleyh . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.5 Apotelèsmata - Telikˆ Sumperˆsmata . . . . . . . . . . . . . . . . . . . . . . 44

4 Paragontopohsh mh arnhtik¸n tanust¸n 45


4.1 Eisagwg  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.2 Teqnikè Anˆlush Upoq¸rwn . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.2.1 Grammikè Teqnikè Anˆlush Upoq¸rwn . . . . . . . . . . . . . . . . 46

4.2.2 Polugrammikè Teqnikè Anˆlush Upoq¸rwn . . . . . . . . . . . . . 46

4.2.3 PARAFAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.2.4 Polugrammik  anˆlush sunistws¸n . . . . . . . . . . . . . . . . . . . 47

4.2.5 Polugrammik  ICA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.3 Paragontopohsh mh arnhtik¸n pinˆkwn . . . . . . . . . . . . . . . . . . . . . 50

4.3.1 Montèlo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4.3.2 Basikì NMF Standard NMF


( ) . . . . . . . . . . . . . . . . . . . . . 51

4.3.3 Topikì NMF Lo alized NMF


( ) . . . . . . . . . . . . . . . . . . . . . 52

4.3.4 Araiì NMF Sparse NMF


( ) . . . . . . . . . . . . . . . . . . . . . . . 53

4.3.5 Taxinìmhsh NMF qwr epbleyh . . . . . . . . . . . . . . . . . . . . 54

4.3.6 Taxinìmhsh NMF me epbleyh . . . . . . . . . . . . . . . . . . . . . . 54

4.3.7 Epektˆsei mejìdwn NMF . . . . . . . . . . . . . . . . . . . . . . . . 56

4.4 Proteinìmene Mèjodoi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.5 NTF Qrhsimopoi¸nta Apoklsei Bregman . . . . . . . . . . . . . . . . . . 64

4.5.1 Prìblhma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.5.2 Bohjhtik  Sunˆrthsh gia NTF . . . . . . . . . . . . . . . . . . . . . 66

4.5.3 Apìdeixh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.5.4 Elaqistopoi¸nta thn bohjhtik  sunˆrthsh . . . . . . . . . . . . . . 67

4.5.5 Eidik  perptwsh: diaqwrsimh () . . . . . . . . . . . . . . . . . . . . 68

4.6 NTF gia Sugkekrimène Apoklsei Bregman . . . . . . . . . . . . . . . . . . 69

4.6.1 NTF me apìklish Kullba k - Leibler . . . . . . . . . . . . . . . . . . . 69

4.6.2 NTF me nìrma Frobenius deutèrou bajmoÔ . . . . . . . . . . . . . . . 69

4.6.3 NTF me apìstash Itakura - Saito . . . . . . . . . . . . . . . . . . . . 69

vi
PERIEQŸ
OMENA PERIEQŸ
OMENA

4.7 Algìrijmoi NTF gia Tanustè 3 Diastˆsewn . . . . . . . . . . . . . . . . . 69

4.7.1 NTF 3h tˆxh me apìklish Kullba k - Leibler . . . . . . . . . . . . . 69

4.7.2 NTF 3h tˆxh me nìrma Frobenius 2ou bajmoÔ . . . . . . . . . . . . 70

4.7.3 NTF 3h tˆxh me apìstash Itakura - Saito . . . . . . . . . . . . . . 70

4.8 Taxinìmhsh NTF me Epbleyh . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.9 Parathr sei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5 Exagwg  Qarakthristik¸n 75
5.1 Eisagwg  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.2 To Prìtupo MPEG-7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5.2.1 Eisagwg  sto MPEG-7 Audio . . . . . . . . . . . . . . . . . . . . . . 76

5.2.2 Perigrafe QamhloÔ Epipèdou . . . . . . . . . . . . . . . . . . . . . . 78

5.3 QrhsimopoioÔmena Qarakthristikˆ . . . . . . . . . . . . . . . . . . . . . . . . 79

5.3.1 Qarakthristikˆ enèrgeia . . . . . . . . . . . . . . . . . . . . . . . . 79

5.3.2 Fasmatikˆ qarakthristikˆ . . . . . . . . . . . . . . . . . . . . . . . . 80

5.3.3 Qronikˆ qarakthristikˆ . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.3.4 Antilambanìmena qarakthristikˆ . . . . . . . . . . . . . . . . . . . . 84

5.3.5 Armonikˆ qarakthristikˆ . . . . . . . . . . . . . . . . . . . . . . . . . 85

6 Peirˆmata 87
6.1 Eisagwg  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

6.2 Dhmiourga Tanust  Dedomènwn . . . . . . . . . . . . . . . . . . . . . . . . . 87

6.3 Epilog  Qarakthristik¸n . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

6.4 Epiprìsjetoi Taxinomhtè . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

6.4.1 Polustrwmatiko Per eptrons . . . . . . . . . . . . . . . . . . . . . . 90

6.4.2 Mhqanè Edrawn Dianusmˆtwn . . . . . . . . . . . . . . . . . . . . . 92

6.5 Peiramatikˆ Apotelèsmata . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

6.6 Mellontikè KateujÔnsei . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

vii
PERIEQŸ
OMENA PERIEQŸ
OMENA

viii
Kefˆlaio 1

Eisagwg 

1.1 Stìqoi diatrib 

Stìqo th diatrib  enai h melèth tou pedou th anang¸rish mousikoÔ edou se polugram-

mikì peribˆllon. Ta teleutaa qrìnia parathretai h dhmiourga megˆlwn bˆsewn dedomènwn

apì mousikˆ kommˆtia kai kajstatai epitaktik  h anˆgkh orgˆnwsh twn dedomènwn. En¸

èqoun anaptuqje ergalea gia thn qeiroknhth orgˆnwsh twn arqewn, o ìgko tou enai

tìso megˆlo pou kajistˆ to egqerhma sqedìn adÔnato. 'Ara, enai epitaktik  h anˆgkh

anˆptuxh ergalewn autìmath taxinìmhsh , anaz thsh , kai anˆkthsh twn arqewn. To

prìblhma th anagn¸rish mousikoÔ edou ìmw den enai tetrimmèno: enai qarakthristikì

ìti den èqei protaje kama genik  taxinìmhsh gia mousikˆ edh kai akìma kai h anjr¸pinh

akrbeia taxinìmhsh ftˆnei to 72% gia mh eidikoÔ sto pedo.

Sthn bibliografa èqoun protaje pollè teqnikè anagn¸rish mousikoÔ edou , pou sti

perissìtere peript¸sei qrhsimopoioÔn kˆpoio algìrijmo mhqanik  mˆjhsh gia na ektelè-

sei thn taxinìmhsh twn arqewn katìpin ekpadeush . Ma ˆllh logik  enai h omadopohsh

twn arqewn me bˆsh ta qarakthristikˆ perigraf  tou , afoÔ oÔtw   ˆllw den upˆrqei

kˆpoia koin¸ apodekt  taxinìmhsh twn mousik¸n eid¸n.

Sthn sugkekrimènh diatrib , ja antimetwpiste to prìblhma th autìmath taxinìmhsh

mousikoÔ edou , qrhsimopoi¸nta ma nèa logik : ìti to s ma tou  qou den enai monodiˆsta-

to, allˆ mpore na anaparastaje me ma poludiˆstath dom , ekmetalleÔonta sto èpakro thn

plhrofora pou upˆrqei sto s ma, kai dnonta èmfash sthn allag  twn qarakthristik¸n

pou sunteloÔntai ston qrìno. H lÔsh sto prìblhma th poludiˆstath anaparˆstash enì

s mato dnetai apì to pedo th polugrammik  ˆlgebra (multilinear algebra ) kai pio sugke-

krimèna apì ti poludiˆstate domè pou onomˆzontai tanustè ( tensors). Katìpin melèth

tou pedou th anˆlush tanust¸n, protenetai sthn diatrib  ma teqnik  gia thn anˆlush kai

taxinìmhsh poludiˆstatwn shmˆtwn. H teqnik  onomˆzetai paragontopohsh mh arnhtik¸n

1
KEFŸ
ALAIO 1. EISAGWGŸ
H

tanust¸n ( non-negative tensor fa torization ) kai baszetai sthn idèa ìti èna opoiod pote

poludiˆstato s ma enai sunduasmì basik¸n shmˆtwn pou den enai grammik¸ exarthmèna.

Parìmoioi algìrijmoi èqoun protaje sth bibliografa, kurw sto pedo th epexergasa

eikìna , allˆ sthn paroÔsa diatrib  dnetai èmfash sth dhmiourga enì algorjmou pou na

efarmìzetai se tanustè ìlwn twn diastˆsewn kai qrhsimopoi¸nta ma omˆda apì mètra

omoiìthta , mh periorzonta thn efarmog  tou algorjmou se èna jèma kai mìno.

AkoloÔjw , gia thn eplush tou probl mato dhmiourgoÔntai tanustè dedomènwn kai e-

farmìzontai oi proteinìmene teqnikè . Exˆgontai qarakthristikˆ perigraf  pou kalÔptoun

fasmatikˆ, qronikˆ, energeiakˆ, antilambanìmena, kai qarakthristikˆ perigraf  mousikoÔ

tìnou. Gia plhrìthta, sta peirˆmata taxinìmhsh qrhsimopoioÔntai kai sun jei algìrij-

moi mhqanik  mˆjhsh , kaj¸ kai algìrijmoi paragontopohsh mh arnhtik¸n pinˆkwn, pou

leitourgoÔn se grammikì peribˆllon. Ta apotelèsmata deqnoun thn uperoq  twn polugram-

mik¸n algorjmwn se sqèsh me tou grammikoÔ algìrijmou .

1.2 Dom  diatrib 

Akolouje h orgˆnwsh th diplwmatik  se kefˆlaia:

 Sto Kefˆlaio 2 dnetai ma eisagwg  sti basikè ènnoie th grammik  kai polugram-

mik  ˆlgebra . Dnetai èmfash ston orismì twn tanust¸n kai sti idiìthtè tou .

Epsh dnetai orismì twn apoklsewn Bregman , oi opoe apoteloÔn thn bˆsh twn

proteinìmenwn algorjmwn.

 Ma sunoptik  diereÔnhsh tou pedou th autìmath anagn¸rish mousikoÔ edou d-

netai sto Kefˆlaio 3. Anafèrontai ta basikˆ dedomèna pou qrhsimopoioÔntai se pei-

rˆmata, ta qarakthristikˆ perigraf  twn dedomènwn mousikoÔ edou , kai oi sun jei

algìrijmoi taxinìmhsh .

 Sto Kefˆlaio 4 dnetai o proteinìmeno algìrijmo paragontopohsh mh arnhtik¸n ta-

nust¸n. Melet¸ntai proteinìmenoi polugrammiko algìrijmoi taxinìmhsh kai anˆlush

dedomènwn sthn bibliografa. Katìpin, meletˆtai to prìblhma th paragontopohsh

mh arnhtik¸n pinˆkwn, kai protenetai o algìrijmo paragontopohsh mh arnhtik¸n ta-

nust¸n, pou apotele genkeush tou prohgoumènou. Tèlo , protenetai èna taxinomht 

gia efarmog  se algorjmou paragontopohsh mh arnhtik¸n tanust¸n.

 Ta qarakthristikˆ prigraf  twn dedomènwn pou qrhsimopoi jhkan sthn diatrib  d-

nontai sto Kefˆlaio 5. Meletˆtai sunoptikˆ to prwtìkollo MPEG-7 , apì to opo-

o qrhsimopoi jhkan kˆpoia qarakthristikˆ. AkoloÔjw , perigrˆfontai analutikˆ ta

qrhsimopoioÔmena qarakthristikˆ anˆloga me thn kathgora sthn opoa an koun.

2
1.2. DOMŸ
H DIATRIBŸ
HS

 Tèlo , sto Kefˆlaio 6 dnetai h peiramatik  diadikasa kai ta apotelèsmata. Parou-

siˆzetai h diadikasa kataskeu  tou tanust  apì ta qarakthristikˆ. Sth sunèqeia,

pragmatopoietai epilog  twn pio katˆllhlwn qarakthristik¸n gia taxinìmhsh. Peri-

grˆfontai sunoptikˆ oi algìrijmoi mhqanik  mˆjhsh pou qrhsimopoi jhkan gia sÔ-

gkrish. Tèlo , dnontai ta peiramatikˆ apotelèsmata maz me sqìlia kai sumperˆsmata.

3
KEFŸ
ALAIO 1. EISAGWGŸ
H

4
Kefˆlaio 2

Logismì Tanust¸n

To parìn kefˆlaio apotele ma eisagwg  sti ènnoie th polugrammik  (  pleiogrammik  )

ˆlgebra . Perièqei basikoÔ orismoÔ kai idiìthte th grammik  ˆlgebra gia lìgou plh-

rìthta kai sÔgkrish . Sthn sunèqeia, dnetai orismì kai basikè idiìthte twn tanust¸n.

Sth sunèqeia gnetai ma parˆjesh statistik¸n mètrwn uyhl  tˆxh , ta opoa perigrˆfo-

ntai me thn qr sh twn tanust¸n. To kefˆlaio klenei me anaforˆ sti apoklsei Bregman ,

pou qrhsimopoioÔntai sthn paroÔsa diatrib , parajètonta ti idiìthtè tou ston tanustikì

q¸ro.

2.1 DianÔsmata kai Pnake

2.1.1 Orismo

Orismì 2.1.1. Dianusmatikì Q¸ro


( V pˆnw se èna pedo F
) Dianusmatikì q¸ro

enai èna sÔnolo stoiqewn, onomazìmena dianÔsmata, ètsi ¸ste, an a; b 2 V kai ; d 2 F :

1.  a + d  b 2 V

2. 9 a: a + ( a) = a a = 0
3. 9 mhdenikì diˆnusma: 0 + a = a

4. 1  a = a; 8a 2 V

5. a + b = b + a

6. a + (b + ) = (a + b) +

7. ( + d)  a =  a + d  a,

 (a + b) =  a +  b,
( d)  a =  (da)
Parˆdeigma 2.1.1. 'Estw a diˆnusma m kou n:
 
a= a a : : : an
1 2
(2.1)

5
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Orismì 2.1.2. Pnaka ( V ; V dianusmatiko q¸roi me diastˆsei , I ; I , antstoi-


) 'Estw 1 2 1 2

qa. JewroÔme 2 dianÔsmata u 2 V ; u 2 V . O q¸ro pou dhmiourgetai apì ìla ta stoiqea


1 1 2 2

th digrammik  apeikìnish u Æ u (u Æ u = Uij ) ston V  V enai o q¸ro dianusmati-


1 2 1 2 1 2

koÔ ginomènou gia tou q¸rou V ; V . 'Ena stoiqeo tou q¸rou dianusmatikoÔ ginomènou
1 2

onomˆzetai pnaka (  m tra) ston V  V . 1 2

'Ena pio praktikì orismì tou pnaka: ma disdiˆstath diˆtaxh stoiqewn enì pedou

F, pou èqei m grammè kai n st le , onomˆzetai pnaka mn sto F. Sumbolzetai w

A = [aij ℄mn , ìpou oi dekte mn orzoun ti dunamikè perioqè twn deikt¸n ij , dhlad 

i = 1; : : : ; m kai j = 1; : : : ; n. Ta dianÔsmata enai 1n pnake . 'Ara sumfwnoÔme ìti ìla

ta dianÔsmata enai dianÔsmata gramm  efex  .

Parˆdeigma 2.1.2. 'Estw A pnaka m  n diastˆsewn:


0 1
a a ::: a n 11 12 1
B C
Ba a : : : a
A=B
B
: : : : : : : :
21

: : : : :
n
22

:
C
C
C
2
(2.2)
 A

am am : : : amn 1 2

2.1.2 Idiìthte

Orismì 2.1.3. Eswterikì Ginìmeno Dianusmˆtwn


( ) An u = (u : : : un) kai v = (v : : : vn ),
1 1

to eswterikì ginìmeno twn u kai v orzetai w :


n
X
uv= ui vi (2.3)

i=1
Ta dianÔsmata u kai v enai orjog¸nia an u  v = 0.
Orismì 2.1.4. Monadiao Pnaka
( ) Orzetai w o tetragwnikì pnaka ( m = n):
I = [Æij ℄nn (2.4)

ìpou Æij Krone ker


to dèlta tou . Gia opoiond pote tetragwnikì pnaka A diastˆsewn nn
isqÔei ìti: IA=AI=A .

Orismì 2.1.5. Nìrma Frobenius 2h Tˆxh


( ) Gia ènan pnaka A diastˆsewn mn
orzetai w :
v
u m n
kAk =
uX X
2
t jaij j2
(2.5)

i=1 j =1

6
2.2. ORISMŸ
OS TANUSTŸ
WN

Orismì 2.1.6. Ginìmeno Pinˆkwn A = [aij ℄mn B = [bij ℄nk


( ) An kai , to ginìmeno twn

A B
pinˆkwn kai AB = [(AB)ij ℄mk
sumbolzetai ij , kai èqei stoiqeo:

n
(AB)ij =
X
air brj (2.6)

r=1

Orismì 2.1.7. Orjog¸nio Tetragwnikì Pnaka


( ) 'Ena tetragwnikì pnaka A=
[aij ℄nn enai orjog¸nio an:

AT = A 1
(2.7)

ìpou AT = [aji℄nn enai o anˆstrofo tou A A kai


1
o antstrofo tou A, dhlad  pnaka

tètoio ¸ste AA = I. 1

2.2 Orismì Tanust¸n

Oi tanustè ( tensors ) apoteloÔn ma genkeush twn dianusmˆtwn kai twn pinˆkwn, ennoi¸n

th grammik  ˆlgebra , kai apoteloÔn basikì antikemeno th Polugrammik  'Algebra . A-

nˆloga me to episthmonikì pedo sto opoo sunant¸ntai, oi tanustè mporoÔn na oristoÔn me

diaforetikoÔ trìpou . Sto majhmatikì pedo th Anˆlush Tanust¸n, orzontai w antike-

mena pou upìkeintai se poludiˆstatou metasqhmatismoÔ , apì èna sÔsthma suntetagmènwn

se èna ˆllo. Oi metasqhmatismo pou qrhsimopoioÔntai onomˆzontai summetablhto ( ova-


riant - apì ton dianusmatikì ston duðkì q¸ro) kai antimetablhto ( ontravariant - apì ton

duðkì q¸ro ston dianusmatikì). Ston klˆdo th fusik  qrhsimopoioÔntai gia na perigrˆ-

youn posìthte ston trisdiˆstato q¸ro. Epsh , o ìro tanust  qrhsimopoietai kai gia

thn perigraf  pedwn. Gia tou parapˆnw lìgou , qrhsimopoioÔntai ekten¸ sto pedo th

mhqanik  suneq¸n mèswn ( ontinuum me hani s ). Perissìtere plhrofore gia to pedo

th polugrammik  ˆlgebra mporoÔn na anazhthjoÔn sto [37℄.

Ston klˆdo twn efarmosmènwn majhmatik¸n, pou den apaitoÔntai metasqhmatismo se

sust mata suntetagmènwn, oi tanustè orzontai w ma genikeumènh grammik  ontìthta pou

mpore na anaparastaje w èna pollaplì ( multiway ) pnaka . Akolouje o orismì twn ta-

nust¸n pou mporoÔn na efarmostoÔn mìno gia grammikoÔ metasqhmatismoÔ suntetagmènwn

( omponent-free approa h ) [51℄. Gia th qr sh tanust¸n se efarmogè pou qrhsimopoioÔntai

w pollaplo pnake , èqei protaje to N-way Toolbox gia peribˆllon MATLAB [2℄. Orismì

twn mikt¸n tanust¸n pou upìkeintai se mh grammikoÔ metasqhmatismoÔ kai apantioÔntai se

kampulìgramme suntetagmène paratjetai sta [12, 48℄. H anˆlus  ma emploutzetai ìmw

me sqìlia sqetikˆ me th sumperiforˆ twn mikt¸n tanust¸n gia lìgou sÔgkrish .

7
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Orismì 2.2.1. Tanust  N -ost  Tˆxh


( V ; V ; : : : ; VN N dianusmatiko q¸roi
) 'Estw 1 2

me diastˆsei I ; I ; : : : ; IN .
1 2 JewroÔme N dianÔsmata u 2 V ; u 2 V ; : : : ; 1 1 2 2

uN 2 VN . O q¸ro pou dhmiourgetai apì ìla ta stoiqea th polugrammik  apeikìnish


u Æ u Æ : : : Æ uN ston V  V  : : :  VN onomˆzetai q¸ro tanust¸n. 'Ena stoiqeo tou
1 2 1 2

q¸rou tanust¸n onomˆzetai tanust  N -ost  tˆxh ston V  V  : : :  VN . 1 2

Gia Vn = R In (n = 1; 2; : : : ; N ) , o q¸ro ginomènou tanust¸n onomˆzetai q¸ro twn

pragmatik¸n ( I1  I  : : :  IN
2 )-tanust¸n, kai sumbolzetai me R I1 I2 :::IN . O antstoiqo

migadikì q¸ro sumbolzetai me C I1 I2 :::IN .

Parˆdeigma 2.2.1. 'Estw tanust  A2R 3 3 3


. JewroÔme ìti aij = 1; aij = 2; aij =
1 2 3

3 (i; j = 1; 2; 3) . O A enai dunatìn na sumboliste qrhsimopoi¸nta pnake :

0 1 0 1 0 1
1 1 1C 2 2 2C 3 3 3C
A = [Aij jAij jAij ℄ Aij Aij Aij
B B B
1 2 1 me 1 =B
1 1 1A
C
2 =B
2 2 2A
C
3 =B
3 3 3A
C (2.8)

1 1 1 2 2 2 3 3 3

Gia lìgou plhrìthta paratjetai o orismì twn tanust¸n me qr sh mh-grammik¸n me-

tasqhmatism¸n ( omponent approa h ):

Orismì 2.2.2. Miktì Tanust  N -ost  Tˆxh


( ) Orzetai A w miktì tanust 

tˆxh ( P + Q), antimetablht  tˆxh P kai summetablht  tˆxh Q , an ta stoiqea tou

ij11ij22:::i
:::jQ
metasqhmatzontai se
P ~ij11ij22:::i
:::jQ
P metabanonta apì to sÔsthma suntetagmènwn xip
(p)

sto sÔsthma suntetagmènwn x~(ipp) gia p = 1; : : : ; P kai apì yj(qq) se y~j(qq) gia q = 1; : : : ; Q me ton
akìloujo kanìna:

:::sQ ys1 ys2


(1) (2)
ysQQ xi xi
( )
xiPP
(1) (2) ( )

~ ij11ij22:::i
:::jQ
= rs11rs22:::r ::: Q ::: P
1 2
(2.9)
P P
 y~j(1)
1
 y~j(2)
2
 y~jQ  x~r  x~r
( )
 x~rP
(1)
1
(2)
2
( )

H exswsh (2.9) enai gnwst  w Kanìna MetasqhmatismoÔ Tanust¸n.

2.3 Prˆxei me Tanustè

Gia tanustè peperasmènh tˆxh kai diastˆsewn orzontai kai oi basikè prˆxei th polu-

grammik  ˆlgebra , pou apoteloÔn genkeush twn basik¸n prˆxewn th grammik  ˆlgebra .

Oi parakˆtw orismo dnontai gia migadikoÔ tanustè , kai profan¸ mporoÔn na efarmostoÔn

kai se tanustè me pragmatikè timè .

8
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

2.3.1 Basikè Prˆxei

Orismì 2.3.1. Prìsjesh Tanust¸n


( ) 'Estw dÔo tanustè A 2 C I I :::IN
1 2
kai B2
C I1 I2 :::IN . To ˆjroismˆ tou (A + B ) 2 C I1 I2 :::IN orzetai w :

(A + B)i i :::iN = ai i :::iN + bi i :::iN


1 2 1 2 1 2
(2.10)

Anˆloga orzetai kai to bajmwtì ginìmeno enì tanust  A me ma pragmatik  stajerˆ ,
pou dhmiourge ènan tanust  dia tˆxh kai diastˆsewn me ton A .

Orismì 2.3.2. Grammikì Sunduasmì


( ) JewroÔme M tanustè A l 2 C I I :::IN
( ) 1 2
,

l = 1; 2; : : : ; M , kai M stajerè  ;  ; : : : ; M 2 R . O tanust 


1 2 B =  A + A +:::+
1
(1)
2
(2)

M A M èqei thn dia tˆxh kai diastˆsei me tou A l (l = 1; 2; : : : ; M ).


( ) ( )

2.3.2 Ginìmena Tanust¸n

Orismì 2.3.3. Eswterikì Ginìmeno


( ) 'Estw tanust  A 2 C I I :::IM 1 2
kai tanust 

B 2 C J J :::JN
1 2
. To eswterikì ginìmeno w pro tou dekte im kai in , me koin  diˆstash
jImj = jJnj = p , orzetai w hA; Bim;n 2 C I I :::Im 1 2 1 Im+1 :::IM J1 J2 :::Jn 1 Jn+1 :::JN :
X
p
(hA; Bim;n)i :::im1 1 im+1 :::iM j1 :::jn 1 jn+1 :::jN = ai :::im
1 1 kim+1 :::iM bj1 :::jn 1 kjn+1 :::jN (2.11)

k=1

Enai dunatìn na oriste eswterikì ginìmeno pˆnw se polloÔ dekte . Epsh , sthn

perptwsh miktoÔ tanust , to eswterikì ginìmeno orzetai mìno anˆmesa se summetablhtoÔ

kai antimetablhtoÔ dekte . Praktikˆ, oi summetablhtè kai antimetablhtè sumperiforè

tou tanust  allhloakur¸nontai, mei¸nonta thn sunolik  tˆxh tou tanust .

Sthn perptwsh tanust¸n 2h tˆxh (pnake ), to eswterikì ginìmeno isoÔtai me to ginì-

meno pinˆkwn (Orismì 2.1.6). Diaisjhtikˆ mpore na ermhneute w to eswterikì ginìmeno

dÔo tanust¸n, oi opooi èqoun ma koin  diˆstash. To apotèlesma enai èna tanust  2h

tˆxh o opoo èqei orzetai sti dÔo mh koinè diastˆsei twn pinˆkwn.

Parˆdeigma 2.3.1. 'Estw tanustè A; B 2 R   2 2 2


. JewroÔme ìti ai i i = 1; bj j j = 2 (1 6
1 2 3 1 2 3

ix ; jx 6 2) . Upologzoume to eswterikì ginìmeno twn A; B me bˆsh thn pr¸th diˆstash w

koin :
2
X
(hA; Bi ; )i i j j =
1 1 2 3 2 3
aki i bkj j
2 3 2 3 (2.12)

k=1
'Ara, gia na upologsoume ta stoiqea tou tanust  hA; Bi ; 2 R   
1 1
2 2 2 2
qrhsimopoioÔme thn

(2.12). Gia to stoiqeo (1,1,1,1): (hA; Bi ; ) 1 1 1111 =a b +a b =2+2=4


111 111 211 211 .

9
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Orismì 2.3.4. Exwterikì Ginìmeno


( ) 'Estw tanust  A 2 C I I :::IM
1 2
kai tanust 

B2 C J1 J2 :::JN . To exwterikì ginìmeno A


B 2 C I1 :::IM J1 :::JN , pou èqei stoiqea twn

tanust¸n A kai B orzetai w :

(A
B)i :::iM j :::jN = ai :::iM bj :::jN
1 1 1 1 (2.13)

'Ara to exwterikì ginìmeno dÔo tanust¸n (to opoo sunantˆtai sthn bibliografa kai w

tanustikì ginìmeno) èqei tˆxh M + N, kai diastˆsei I1  : : :  IM  J  : : :  JN


1 . Sthn

perptwsh pinˆkwn (tanust¸n tˆxh 2) ston dio dianusmatikì q¸ro, to exwterikì ginìmeno

onomˆzetai ginìmeno Krone ker , lìgw tou GermanoÔ majhmatikoÔ Leopold Krone ker . Sthn

perptwsh exwterikoÔ ginomènou mikt¸n tanust¸n, prokÔptei tanust  me summetablht  tˆxh

sh me to ˆjroisma th summetablht  tˆxh twn dÔo kai antstoiqa prokÔptei tanust  me

antimetablht  tˆxh sh me to ˆjroisma th antimetablht  tˆxh twn dÔo.

Parˆdeigma 2.3.2. 'Estw pnake A 2 R I I 1 2


diastˆsewn 22 me aij = 1 kai B 2 R I I 3 4

diwn diastˆsewn bkl = 2. To exwterikì ginìmeno twn dÔo pinˆkwn enai tanust  me stoiqea:

(A
B)i i i i = ai i bi i
1 2 3 4 1 2 3 4
(2.14)

'Ara to prokÔpton exwterikì ginìmeno enai tanust  4h tˆxh diastˆsewn 2222 . To

stoiqeo (1,1,1,1) tou exwterikoÔ ginomènou isoÔtai me: (A


B) 1111 =12=2 .

Sthn perptwsh pou B 2 R I I1 2


, dhlad  oi pnake an koun ston dio dianusmatikì q¸ro,

tìte to exwterikì tou ginìmeno isoÔtai me to ginìmeno Krone ker:


0 1
2 2 2 2
B C
2 2 2 2C
A
B=B
B C (2.15)
2 2 2 2C
B
A

2 2 2 2
S' aut n thn perptwsh to apotèlesma enai pnaka diastˆsewn 2222=44 .

2.3.3 Monadiao kai Isotropikì Tanust 

Orismì 2.3.5. Monadiao Tanust 


( ) O monadiao tanust  I 2 C I :::IN J :::JN
1 1

orzetai w ma polumetablht  genkeush tou Dèlta tou Krone ker kai èqei stoiqea:

N
Y
Ii :::iN j :::jN =
1 1
Æij11:::i
:::jN
N = Æik jk (2.16)

k=1

10
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

Praktikˆ, o monadiao tanust  kataskeuˆzetai apì to exwterikì ginìmeno poll¸n mo-

nadiawn pinˆkwn I . Koin¸ , o monadiao tanust  orzetai gia ˆrtie tˆxei (akolouje

parˆdeigma).

Parˆdeigma 2.3.3. O monadiao tanust  I 2 C I I J J 1 2 1 2


èqei stoiqea:

Ii i j j = Æi j Æi j ;
1 2 1 2 1 1 2 2
(2.17)

dhlad  èqei tim  1 ìtan i = j


1 1 kai i = j
2 2 kai 0 alloÔ. An jewr soume ìti o I èqei

diastˆsei 2222 , tìte parnei tim  1 sta stoiqea: (1,1,1,1), (1,2,1,2), (2,1,2,1),

(2,2,2,2). Antistoqw , parnei tim  0 sta stoiqea (1,1,1,2), (1,1,2,1), (1,2,1,1), (2,1,1,1),

(1,2,2,2), (1,2,2,1), (1,1,2,2), (2,2,1,1), (2,2,2,1), (2,1,1,2), (2,1,2,2), kai (2,2,1,2).

Epeid  ta stoiqea tou monadiaou tanust  den exart¸ntai apì sugkekrimènh epilog  bˆ-

sh , o monadiao tanust  onomˆzetai isotropikì ( isotropi ). Sugkekrimèna, èna tanust 

pou oi timè twn stoiqewn tou enai anallowte me thn peristrof  tou sust mato sunte-

tagmènwn onomˆzetai isotropikì . 'Oloi oi tanustè mhdenik  tˆxh (bajmwtˆ megèjh) enai

isotropiko, en¸ den upˆrqoun isotropikˆ dianÔsmata. O monadikì isotropikì tanust 

deÔterh tˆxh enai to Dèlta tou Krone ker . O monadikì isotropikì tanust  trth

tˆxh enai to sÔmbolo antimetˆjesh ijk :


8
0 >
>
>
<
gia i = j ,   j = k,   k = i
ijk = 1
>
gia (i; j; k) 2 f(1; 2; 3); (2; 3; 1); (3; 1; 2)g (2.18)
>
(i; j; k) 2 f(1; 3; 2); (3; 2; 1); (2; 1; 3)g
>
:
1 gia

o opoo sthn bibliografa anafèretai suqnˆ w tanust  Levi-Civita.

2.3.4 Sustol  Tanust 

Orismì 2.3.6. Sustol  Tanust 


( ) JewroÔme tanust  A 2 C I I :::IN
1 2
. H sustol 

( ontra tion ) tou A w pro tou dekte ip kai iq , me koin  diˆstash u, enai o tanust 

hAip;q , me stoiqea:

u
X
(hAip;q)i :::ip1 1 ip+1 :::iq 1 iq+1 :::iN = ai :::ip
1 1 kip+1 :::iq 1 kiq+1 :::iN (2.19)

k=1

H sustol  enì tanust  mei¸nei thn tˆxh tou katˆ 2. Enai epsh dunatìn na oriste

sustol  gia perissìterou apì 2 dekte . Sthn perptwsh pnaka, h sustol  enai antstoiqh

me thn eÔresh tou qnou ( tra e ) tou pnaka. Sthn perptwsh miktoÔ tanust , h sustol 

orzetai anˆmesa w pro ènan summetablhtì kai ènan antimetablhtì dekth.

11
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Axzei na shmeiwje h sqèsh tou eswterikoÔ ginomènou, tou exwterikoÔ ginomènou kai th

sustol  tanust¸n. To exwterikì ginìmeno 2 tanust¸n A; B kai h metèpeita sustol  tou

ginomènou w pro dedomènou dekte , dnei dio apotèlesma me to eswterikì ginìmeno twn

dÔo tanust¸n w pro tou diou dekte .

Parˆdeigma 2.3.4. JewroÔme tanust  A 2 R I I I I


1 2 3 4
, me diastˆsei 2222 kai

timè ai i i i = 1.
1 2 3 4
H sustol  tou A w pro tou dekte i ;i2 3 enai:

2
X
(hAi ; )i ;i =
2 3 1 4
ai kki = ai
1 4 1 11 4i + ai i
1 22 4
(2.20)

k=1

'Ara to apotèlesma enai o pnaka A0 2 RI I 1 4


me stoiqea:

!
2 2
(hAi ; )i ;i = A0 = (2.21)
2 2
2 3 1 4

2.3.5 Ginìmeno Tanust  me Pnaka

Orismì 2.3.7. Ginìmeno Tanust  me Pnaka


( ) JewroÔme tanust  A 2 C I I :::IN
1 2
kai

pnaka B 2 C Jn In . To ginìmeno tou A ep ton B w pro thn koin  diˆstash In orzetai w :
A n B = hA; Bin; : 2 (2.22)

Koin¸ , to ginìmeno tanust  me pnaka orzetai w to eswterikì ginìmenì tou pˆnw se

koinoÔ dekte . Analutikˆ, ta stoiqea tou ginomènou upologzontai me bˆsh thn parakˆtw

sqèsh:
In
(A n B)i :::in
X
1 1 jn in+1 :::iN = ai :::k:::iN bjn k
1
(2.23)

k=1
ìpou in enai o koinì dekth tou tanust  kai tou pnaka. ParathroÔme ìti to ginìmeno

tanust  me pnaka, o tanust  pou prokÔptei èqei thn dia tˆxh me ton arqikì tanust :

(A n B) 2 C I :::Jn:::IN
1
. Aplˆ o n-ostì dekth an kei plèon sto Jn .
Jewr¸nta tanust  A2 C I1 I2 :::IN kai pnake B 2 C
Jn In kai C 2 C Jm Im orzontai

dÔo idiìthte pou aforoÔn to ginìmeno tanust  me pnake :

(A n B) m C = (A m C) n B = A n B m C (2.24)

An t¸ra jewr soume tanust  C 2 C Kn Jn :


(A n B) n C = A n (C  B) (2.25)

12
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

H exswsh (2.24) ekfrˆzei thn prosetairistik  idiìthta tou pollaplasiasmoÔ tanust  me pna-

ke . To sÔmbolo  sthn (2.25) ekfrˆzei ton pollaplasiasmì dÔo pinˆkwn, ìpw diatup¸jhke

ston orismì (2.1.6). Apì thn (2.25) fanetai h shmasa tou sumbolismoÔ n .

Sthn perptwsh ginomènou 3 pinˆkwn, èstw A2


B 2 C I I , kai C 2 C I I ,
C I1 I2 , 3 1 4 2

to ginìmeno B  A  C
H mpore na grafe w : A  B  C. Fanetai kajarˆ ìti o B
1 2

pollaplasiˆzetai w pro thn I diˆstash, en¸ o C w pro thn I


1 diˆstash. Ston N - 2

diˆstato q¸ro o parapˆnw sumbolismì lÔnei to prìblhma th eÔresh sumbolism¸n gia

genikeumènou anˆstrofou pinˆkwn.

Parˆdeigma 2.3.5. JewroÔme tanust  A 2 R I I I 1 2 3


, me diastˆsei 222 kai timè

ai i i = 1.
1 2 3
Epsh jewroÔme pnaka B2 R I4 I1 , diastˆsewn 22 :

B = 20 12
To ginìmeno A B1 dnetai apì thn parakˆtw sqèsh:

B)i i i =
2
X
(A  1 4 2 3 aki i bi k = a i i bi k + a i i bi k
2 3 4 1 2 3 4 2 2 3 4
k=1

'Ara to ginìmeno parnei ti ex  timè : (A  1 B) i i = 3 (A  B) i i = 2


1 2 3
, 1 2 2 3
.

!
3 3 2 2
A B= 1
3 3 2 2

2.3.6 Bajmwtì Ginìmeno

Orismì 2.3.8. Bajmwtì Ginìmeno Tanust¸n


( ) JewroÔme tanustè A; B 2 C I I :::IN
1 2
.

To bajmwtì ginìmeno ( s alar produ t ) twn A kai B , orzetai w :

I1 X
X I2 IN
X
hA; Bi = ::: ai i :::iN bi i :::iN
1 2 1 2
(2.26)

i1 =1 i2 =1 iN =1

'Opw fanetai apì thn (2.26), to bajmwtì ginìmeno tanust¸n apotele genkeush tou

eswterikoÔ ginomènou dÔo dianusmˆtwn, pou orsthke ston Orismì 2.1.3. Sthn perptwsh

tanust¸n deÔterh tˆxh , èstw A; B 2 C I I 1 2


, isqÔei h akìloujh idiìthta:

I1 X
I2
hA; Bi = ai i bi i = tr(BH A)
X
1 2
(2.27)
1 2
i1 =1 i2 =1

13
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

ìpou tr() sumbolzei to qno enì pnaka, dhlad  to ˆjroisma twn stoiqewn th kÔria

diagwnou tou. ParathroÔme ìti to bajmwtì ginìmeno dÔo tanust¸n isoÔtai me to eswterikì

tou ginìmeno, w pro ìlou tou (koinoÔ ) dekte . Me bˆsh ton orismì tou bajmwtoÔ

ginomènou, mpore na oriste kai h genkeush th nìrma Frobenius gia tanustè .

Orismì 2.3.9. Nìrma Frobenius Tanust¸n 2ou BajmoÔ


( ) JewroÔme tanust  A2
C I1 I2 :::IN . H nìrma Frobenius 2ou bajmoÔ tou A orzetai w :

p
kAk = hA; Ai
2 (2.28)

H nìrma Frobenius mpore na qrhsimopoihje san dekth gia thn mètrhsh tou megèjou

enì tanust . To tetrˆgwno th nìrma ekfrˆzei thn enèrgeia tou tanust .

2.3.7 Orjogwniìthta Tanust¸n

O parakˆtw orismì twn amoibaa orjogwnwn ( mutually orthogonal ) tanust¸n apotele ge-

nkeush tou antstoiqou orismoÔ gia dianÔsmata.

Orismì 2.3.10. Amoibaa Orjog¸nioi Tanustè


( ) JewroÔme tanustè A; B 2 C I I IN
1 2
.

Oi A; B enai amoibaa orjog¸nioi ( A?B ) an isqÔei:

hA; Bi = 0 (2.29)

Sthn sunèqeia, ja perigrafoÔn ènnoie th orjogwniìthta gia ma sugkekrimènh omˆda

tanust¸n, tou aposuntejeimènou ( de omposed ) tanustè .

Orismì 2.3.11. Aposuntejeimèno Tanust 


( ) JewroÔme tanust  A 2 C I I IN
1 2
.

O A onomˆzetai aposuntejeimèno , an mpore na grafe sth morf :

A = a
a

a N
(1) (2) ( )
(2.30)

ìpou ai
( )
dianÔsmata me ai ( )
2 C Ii ; i = 1; 2; : : : ; N
kai sumbolzei to ginìmeno Krone ker .

Parˆdeigma 2.3.6. JewroÔme tanust  A 2 R I I I 1 2 3


, me A = (A ij jA ij jA ij )
1 2 3 , ìpou:

0 1 0 1 0 1
0 0 0C 1 0 1C 0 0 0C
A ij = 0 0 0C
1
B
B
 A
A ij =
2
B
B

0 0 0CA
A ij = 0 0 0C
3
B
B
 A
0 0 0 1 0 1 0 0 0
O A enai aposuntejeimèno , kaj¸ mpore na grafe w to exwterikì ginìmeno tri¸n dianu-

smˆtwn: a = (1; 0; 1); a = (0; 1; 0); a = ( 1; 0; 1), me A = a


(1) (2) (3) (1)

a
a(2) (3)
.

14
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

Orismì 2.3.12. Pl rw Orjog¸nioi Tanustè


( ) JewroÔme tanustè A; B 2 C I I IN
1 2

oi opooi mporoÔn na aposuntejoÔn se ginìmena Krone ker dianusmˆtwn (aposuntejeimènoi

tanustè ). Oi A; B enai pl rw orjog¸nioi ( ompletely orthogonal ), an:

a i ?b i
( ) ( )
, a i  b i = 0; 8i = 1; : : : ; N:
( ) ( )
(2.31)

H sqèsh twn A kai B sumbolzetai w : A? B .

Parˆdeigma 2.3.7. JewroÔme ton tanust  A tou Paradegmato 2.3.6, kaj¸ kai ton apo-

suntejeimèno tanust  B 2 R I I I 1 2 3
. O B èqei san bˆsh ta dianÔsmata b = (0; 1; 0); a =
(1) (2)

(1; 0; 1); a (3)


= (0; 1; 0) . ParathroÔme ìti a (1)
?b ; a ?b ; a ?b
(1) (2) (2) (3) (3)
. 'Ara prokÔptei to

sumpèrasma ìti oi A kai B enai pl rw orjog¸nioi.

Orismì 2.3.13. Isqurˆ Orjog¸nioi Tanustè


( ) JewroÔme tanustè A; B 2 C I I IN
1 2

oi opooi mporoÔn na aposuntejoÔn se ginìmena Krone ker dianusmˆtwn (aposuntejeimènoi

tanustè ). Oi A; B enai isqurˆ orjog¸nioi ( strongly orthogonal ), an:

a i ?b i
( ) ( )
  a i = b i ;
( ) ( )
8i = 1; : : : ; N: (2.32)

H sqèsh twn A kai B sumbolzetai w : A?sB .

Parˆdeigma 2.3.8. JewroÔme ton tanust  A tou Paradegmato 2.3.6, kaj¸ kai ton apo-

suntejeimèno tanust  B2 R I1 I2 I3 . O B èqei san bˆsh ta dianÔsmata b = (0; 1; 0); a =
(1) (2)

(0; 1; 0); a = ( 1; 0; 1)
(3)
. ParathroÔme ìti a (1)
?b ; a
(1) (2)
= b ; a = b . 'Ara prokÔ-
(2) (3) (3)

ptei to sumpèrasma ìti oi A kai B enai isqurˆ orjog¸nioi.

Me bˆsh tou OrismoÔ 2.3.10-2.3.13, prokÔptei h parakˆtw idiìthta gia dÔo aposunte-

jeimènou tanustè A; B 2 C I I IN 1 2


:

A? B ) A?sB ) A?B (2.33)

O parakˆtw orismì perigrˆfei thn ènnoia tou upo-tanust , pou qrhsimopoietai ston

orismì enì orjog¸niou tanust . Se antistoiqa me ton orismì enì orjog¸niou pnaka

(tou opoou ìla ta dianÔsmata pou ton apoteloÔn enai orjog¸nia anˆ dÔo), h sunj kh pou

qrhsimopoietai enai oi upo-tanustè (antstoiqo me tou upopnake ) na enai orjog¸nioi anˆ

dÔo.

Orismì 2.3.14. Upo-tanust 


( ) O upo-tanust  ( subtensor Ain  2 C I In
) =
1 1 In+1 IN
enì tanust  A 2 C I I IN
1 2
dhmiourgetai krat¸nta mìno thn -ost  tim  ston dekth

in .

15
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Orismì 2.3.15. 'Olo-Orjog¸nio Tanust 


( ) JewroÔme tanust  A 2 C I I IN
1 2
. O

A enai ìlo-orjog¸nio ( all-orthogonal ), an gia kˆje ; , me th sunj kh =6  , isqÔei ìti:

hAin  ; Ain  i = 0
= = (2.34)

Sto [49℄ dnontai oi orismo twn orjogwnwn tanust¸n, maz me basikè idiìthtè tou . E-

psh , protenontai algìrijmoi gia thn eÔresh orjog¸nia aposÔnjesh tanust¸n, dhlad  thn

anaparˆstash enì tanust  w ˆjroisma orjogwnwn tanust¸n (perissìtera sthn Enìthta

2.3.9).

2.3.8 Anˆptugma Tanust 

Pollˆ probl mata sthn anˆlush tanust¸n enai dunatìn na epilujoÔn topojet¸nta ti ti-

mè twn tanust¸n se pnake antstoiqou megèjou . Basikì lìgo enai ìti h anaparˆstash

enì tanust  se ènan pnaka enai pio katanoht  apì programmatistik  allˆ kai diaisjhtik 

ˆpoyh. Epsh , pollè gl¸sse programmatismoÔ kai pakèta anˆlush dedomènwn den upo-

sthrzoun poludiˆstate domè . Pollè apì ti idiìthte th polugrammik  ˆlgebra èqoun

antistoiqe ston disdiˆstato q¸ro th grammik  ˆlgebra , metatrèponta thn eplush e-

nì probl mato ston tanustikì q¸ro se eplush poll¸n upo-problhmˆtwn ston q¸ro twn

pinˆkwn.

Orismì 2.3.16. Anˆptugma Tanust 


( ) JewroÔme tanust  A 2 C I I IN
1 2
. To anˆ-

ptugma tou A w pro ton n-ostì dekth enai o pnaka An ( ) 2 C In  In In :::IN I I :::In
( +1 +2 1 2 1)
.

To stoiqeo ai i :::iN
1 2
tou A brsketai sthn in seirˆ tou A n , kai sthn st lh:
( )

(in 1)In In : : : IN I I : : : In + (in


+1 +2 +3 1)In In : : : IN I I : : : In + : : :
1 2 1 +2 +3 +4 1 2 1

+(iN 1)I I : : : In + (i 1)I I : : : In + (i 1)I I : : : In + : : : + in :


1 2 1 1 2 3 1 2 3 4 1 1

Parˆdeigma 2.3.9. JewroÔme ton tanust  A 2 C I I I


1 2 3
. O A èqei timè aij = 1; aij =
1 2

2; aij = 3 (i; j = 1; 2; 3)
3 .

To anˆptugma A (1) tou A enai pnaka diastˆsewn I 1 I I =333= 39


2 2 :

0 1
1 2 3 1 2 3 1 2 3C
A (1)
B
=B
 1 2 3 1 2 3 1 2 3 A
C

1 2 3 1 2 3 1 2 3
16
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

To anˆptugma A (2) tou A enai pnaka diastˆsewn I2 I I =333= 39


3 1 :

0 1
1 1 1 2 2 2 3 3 3C
A = (2)
B
B
 1 1 1 2 2 2 3 3 3CA
1 1 1 2 2 2 3 3 3
To anˆptugma A (3) tou A enai pnaka diastˆsewn I3 I I =333= 39
1 2 :

0 1
1 1 1 1 1 1 1 1 1C
A (3)
B
=B

2 2 2 2 2 2 2 2 2CA
3 3 3 3 3 3 3 3 3

Ma polÔ qr simh idiìthta tou anaptÔgmato tanust  enai sthn perptwsh aposÔnjesh

tanust  ( tensor de omposition ). Sthn aposÔnjesh, èna tanust  A 2 C I I IN 1 2


enai

dunatìn na aposunteje sto ginìmeno enì tanust  B 2 C J J JN


1 2
ep N pnake :

A = B  C  C    N C N
1
(1)
2
(2) ( )
(2.35)

ìpou oi pnake C i 2 C Ji Ii .


( )
An ekfraste h sqèsh (2.35) se anˆptugma tanust  w pro

ton n-ostì dekth:

An
( ) = C n  B n  [C
C
  
C n
( )
( )
(1) (2) ( 1)

C n
  
C N ℄T
( +1) ( )
(2.36)

ìpou
sumbolzei to ginìmeno Krone ker .

2.3.9 Bajmì Tanust 

Sto pedo th grammik  ˆlgebra , o grammobajmì   apl¸ bajmì ( rank ) enì pnaka

isoÔtai me ton arijmì twn grammik¸ anexˆrthtwn gramm¸n   sthl¸n tou pnaka. 'Ena

enallaktikì orismì tou bajmoÔ enai o arijmì twn mh mhdenik¸n oriak¸n tim¸n (idiazous¸n

tim¸n - singular values ) tou pnaka. Sto pedo th polugrammik  ˆlgebra , h ènnoia tou

bajmoÔ genikeÔetai, anˆloga me th morf  tou tanust . Sthn perptwsh miktoÔ tanust ,

lìgw twn periorism¸n pou efarmìzontai sthn kataskeu  tou, o bajmì isoÔtai me ton arijmì

twn summetablht¸n kai antimetablht¸n deikt¸n. Sthn perptwsh tanust  N -ost  tˆxh

( omponent-free approa h ), orzontai pollè diaforetikè ènnoie gia to bajmì tanust .

Orismì 2.3.17. Tanust  1ou BajmoÔ


( ) JewroÔme tanust  A 2 C I I IN
1 2
. An o

A mpore na grafe w exwterikì ginìmeno N dianusmˆtwn (enai aposuntejeimèno ), tìte o

bajmì tou isoÔtai me 1 ( Rank-1 tensor ).

17
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

O parapˆnw orismì brsketai se apìluth antistoiqa me ton orismì enì pnaka me bajmì

1 sthn grammik  ˆlgebra: o pnaka pou dhmiourgetai apì to exwterikì (kartesianì) ginìmeno

dÔo dianusmˆtwn èqei bajmì so me 1.

Orismì 2.3.18. Bajmì Tanust 


( ) JewroÔme tanust  A 2 C I I IN
1 2
. 'Estw ìti o A
ikanopoie apì thn sqèsh:
r
X
A= i  U i ( )
(2.37)

i=1
ìpou i > 0; i = 1; : : : ; r stajerè kai U i ( )
2 C I1 I2 IN tanustè me bajmì 1. O bajmì tou

A , o opoo sumbolzetai me rank(A) enai to elˆqisto r ètsi ¸ste o A mpore na ekfraste


sÔmfwna me thn (2.37).

Orismì 2.3.19. Orjog¸nio Bajmì Tanust 


( ) JewroÔme tanust  A 2 C I I IN
1 2
.

'Estw ìti o A ikanopoie thn sqèsh (2.37). O orjog¸nio bajmì ( orthogonal rank A
) tou ,

o opoo sumbolzetai me rank? (A), enai to elˆqisto r ètsi ¸ste o A ikanopoie thn (2.37)

kai isqÔei ìti U ?U ; 8i 6= j


i
( ) j
( )
. H parapˆnw aposÔnjesh onomˆzetai orjog¸nia aposÔnjesh

bajmoÔ.

Orismì 2.3.20. Isqurˆ Orjog¸nio Bajmì Tanust 


( ) JewroÔme tanust  A2
C I1 I2 IN . 'Estw ìti o A ikanopoie thn sqèsh (2.37). O isqurˆ orjog¸nio bajmì

( strong orthogonal rank ) touA , o opoo sumbolzetai me rank?s (A), enai to elˆqisto r ètsi
¸ste o A ikanopoie thn (2.37) kai isqÔei ìti U i ?sU j ; 8i 6= j
( ) ( )
.

Axzei na shmeiwje ìti, sthn perptwsh tanust¸n deÔterh tˆxh , h aposÔnjesh bajmoÔ,

h orjog¸nia aposÔnjesh bajmoÔ kai h isqurˆ orjog¸nia aposÔnjesh bajmoÔ tautzontai me

thn aposÔnjesh oriak¸n tim¸n tou pnaka (perissìtera sto [49℄).

Orismì 2.3.21. DianÔsmata n-ostoÔ Dekth


( ) JewroÔme tanust  A 2 C I I IN
1 2
.

Ta dianÔsmata n-ostoÔ dekth tou A èqoun diˆstash In kai prokÔptoun èqonta eleÔjero

ton dekth in (1 6 in 6 In ) kai krat¸nta tou upìloipou dekte stajeroÔ .

Orismì 2.3.22. n-ostì Bajmì Tanust 


( ) JewroÔme tanust  A 2 C I I IN
1 2
. O

n-ostì bajmì tou A sumbolzetai me rankn(A) kai enai h diˆstash tou dianusmatikoÔ

q¸rou pou dhmiourgetai apì ta dianÔsmata n-tˆxh tou A.


O orismì tou dianÔsmato n-ostoÔ dekth apotele genkeush twn dianusmˆtwn gram-

m  kai dianusmˆtwn st lh enì pnaka. Antstoiqa, o n-ostì bajmì tanust  apotele

genkeush tou grammobajmoÔ. Sthn perptwsh tou tanust  ìmw , oi n-osto bajmo den tau-
tzontai metaxÔ tou aparathta. Kai sthn perptwsh pou tautzontai, upˆrqei perptwsh na

diafèroun apì ton bajmì tou tanust . H sqèsh pou isqÔei metaxÔ tou bajmoÔ tanust  kai

tou n-ostoÔ bajmoÔ tanust  enai: rankn (A) 6 rank(A).

18
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

Oi n-osto bajmo enì tanust  mporoÔn na upologistoÔn eÔkola qrhsimopoi¸nta to

antstoiqo anˆptugma tou tanust :

rankn (A) = rank( An) ( ) (2.38)

H eÔresh tou bajmoÔ enì tanust  enai polÔ pio sÔnjeto se sqèsh me thn eÔresh th

n-ostoÔ bajmoÔ, giat jewre ton tanust  san ma n-diˆstath posìthta. O kajorismì tou
bajmoÔ enì I  I    IN tanust  enai akìma anoiktì prìblhma sthn bibliografa. 'Eqoun
1 2

protaje teqnikè gia thn eÔresh bajmoÔ tanust¸n sugkekrimènh tˆxh kai diastˆsewn (pq.

gia tanustè 2  2    2 ). Epsh , apodeiknÔetai ìti to prìblhma eÔresh tou bajmoÔ

enì tanust  3h tˆxh enai ekjetik  poluplokìthta . Tèlo , apodeiknÔetai ìti h tim 

tou bajmoÔ tou tanust  exartˆtai apì to pedo sto opoo orzetai o tanust  . Koin¸ , o

bajmì enì pragmatikoÔ tanust  pou orzetai sto pedo twn pragmatik¸n arijm¸n den èqei

dia tim  me ton bajmì tou diou tanust  an oriste sto migadikì pedo. Me thn parapˆnw

parat rhsh fanetai h diaforˆ sthn eÔresh tou bajmoÔ enì tanust  se sqèsh me thn eÔresh

tou bajmoÔ enì pnaka, ìpou èqoun protaje pollè teqnikè gia thn eÔres  th (anˆlush

oriak¸n tim¸n, aposÔnjesh QR , apaloif  Gauss ktl).

Parˆdeigma 2.3.10. JewroÔme ton tanust  A 2 C I I I


1 2 3
tou Paradegmato 2.3.6. Ef>

ìson A=a
a
a
(1) (2) (3)
, o bajmì tou A den mpore na enai megalÔtero apì 1, sÔmfwna

me thn (2.37). Epeid  profan¸ o bajmì enì tanust  enai jetikì akèraio , o bajmì tou

A isoÔtai me 1.

Parˆdeigma 2.3.11. JewroÔme ton tanust  A 2 C I I 1 2


. O A èqei timè a11 j =a 22j =
1; a j = a j = 2 (j = 1; 2)
12 21 .

To anˆptugma A (1) tou A enai o 22 pnaka :

!
1 1 2 2
A =(1)
2 2 1 1
O 1-ostì bajmì tou A isoÔtai me ton bajmì tou A (1) : rank 1 (A) = 2 .

To anˆptugma A (2) tou A enai o 22 pnaka :

!
1 2 1 2
A =(2)
2 1 2 1
O 2-ostì bajmì tou A isoÔtai me ton bajmì tou A (2) : rank 2 (A) = 2.

To anˆptugma A (3) tou A enai o 22 pnaka :

!
1 2 2 1
A =(3)
1 2 2 1
19
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

O 3 -ostì bajmì tou A (A) = 1


isoÔtai me ton bajmì tou A (3) : rank 3 . Ja upologsoume

t¸ra ton bajmì tou A A = a


a
(a + a ) + a
a
(a + a )
: isqÔei ìti
(1) (2) (1) (3) (3) (2) (1) (3)
,

ìpou a = (0 1) a = (2 1)
(1)
, a = (1 0)
(2)
A
, kai
(3)
. 'Ara, o bajmì tou den enai megalÔtero

apì 2. Efoson rank (A) = rank (A) = 2


1 A 2 , o bajmì tou isoÔtai me 2.

2.3.10 Uper-summetriko Tanustè

Sto pedo th grammik  ˆlgebra , h ènnoia th summetra sunantˆtai se dÔo ekdoqè : thn

pragmatik  kai thn ermitian  summetra. Sto pedo th polugrammik  ˆlgebra , h pragmati-

k  kai ermitian  summetra genikeÔontai sthn ènnoia th uper-summetra ( super-symmetry ).

Protenetai epsh kai h ènnoia th summetra anˆ dÔo ( pairwise symmetry ), pou apotele

ma ligìtero perioristik  genkeush th summetra .

Jewr¸nta ènan tanust  A2 R I I I , h ènnoia th pragmatik  summetra mpore

na genikeute jewr¸nta ìti ta stoiqea tou tanust  ai i :::iN èqoun dia


1 2
tim  me stoiqea

pou prokÔptoun apì opoiad pote antimetˆjesh twn deikt¸n i ; i ; : : : ; iN .


1 2 Sthn perptwsh

migadikoÔ tanust , h genkeush den enai tetrimmènh, kai sqetzetai me thn anaparˆstash

omogen¸n polu¸numwn ston tanustikì q¸ro.

Orismì 2.3.23. Sqetizìmeno Polu¸numo Tanust 


( A 2 R I I I
) JewroÔme tanust 

kai diˆnusma x 2 RI . To sqetizìmeno polu¸numo ( asso iated polynomial A(x) A ) tou or-

zetai w :

I X
X I I
X
A(x) = A  x  x : : : N x =
1 2 ::: ai i :::iN xi xi : : : xiN
1 2 1 2
(2.39)

i1 =1 i2 =1 iN =1

Sthn perptwsh migadikoÔ tanust  A 2 C I I I , to sqetizìmeno polu¸numo orzetai w :

I X
X I I
X
A(x) = A  x~  x~ : : : N x~ N =
1
(1)
2
(2) ( )
::: ai i :::iN x~i x~i : : : x~iNN
1 2
(1)

1
(2)

2
( )
(2.40)

i1 =1 i2 =1 iN =1

ìpou ta dianÔsmata x~ n
( )
isoÔntai ete me to xn ( )
ete me to x n (n = 1; 2; : : : ; N ).
( )

Orismì 2.3.24. Uper-summetrikì Tanust 


( ) JewroÔme tanust  A 2 C I I I kai

to sqetizìmeno polu¸numo A(x) . O A enai uper-summetrikì an: 1) Gia kˆje antimetˆjesh P


ètsi ¸ste x~P(i1 ) x~P(i2 ) : : : x~PN(iN) )
(1) (2) (
= x~i x~i : : : x~iN
(1)

1
(2)

2
( N)
aP i i :::iN = ai i :::iN
, isqÔei ìti ( 1 2 ) 1 2
. 2) Gia kˆje

antimetˆjesh P ètsi ¸ste x~P i x~P i : : : x~P iN = (~xi x~i : : : x~iNN )


(1)

( 1)
(2)

(
N
2)
(

(
)

)
aP i i :::iN =
(1)

1
(2)

2
( )
isqÔei ìti ( 1 2 )

ai i :::iN .
1 2

'Oson aforˆ miktoÔ tanustè , orzetai h orizìntia summetra ( horizontal symmetry ) w

antallag  dÔo summetablht¸n   antimetablht¸n deikt¸n tou tanust . Antstoiqa, h kˆjeth

summetra ( verti al symmetry ) orzetai w antallag  anˆmesa se ènan summetablhtì kai

20
2.3. PRŸ
AXEIS ME TANUSTŸ
ES

ènan antimetablhtì dekth. H uper-summetra se miktoÔ tanustè orzetai w sunduasmì

orizìntia kai kˆjeth summetra . Se ènan uper-summetrikì tanust  me p summetablhtoÔ

kai p antimetablhtoÔ dekte , prèpei na isqÔoun 2  (p!) 2


orizìntie kai kˆjete summetre .

Axzei na shmeiwje ìti to sqetikì polu¸numo enì uper-summetrikoÔ tanust  parnei mìno

pragmatikè timè . Epsh , o orismì th uper-summetra enai anexˆrthto apì thn bˆsh

sthn opoa orzetai to diˆnusma x. Oi uper-summetriko tanustè èqoun megˆlh qr sh sto

pedo th statistik  uyhl  tˆxh ( higher-order statisti s ), pou ja perigrafe sunoptikˆ se

proseq  enìthta. Akolouje èna pio qalarì orismì th summetra tanust¸n, h summetra

tanust  anˆ dÔo.

Orismì 2.3.25. Summetrikì Tanust  Anˆ DÔo


( ) JewroÔme tanust  A 2 C I I I .

O A enai summetrikì anˆ dÔo an, gia kˆje zeÔgo deikt¸n ( in ; in ), upˆrqei ma antimetˆjesh
1 2

Pn1 n2 ètsi ¸ste aPn n i i :::iN )


1 2( 1 2
= ai i :::iN aPn n
1 2   i i :::iN )
1 2( 1 2
= ai i :::iN .
1 2

Parˆdeigma 2.3.12. JewroÔme ton tanust  A2R  2 2 2


. O A parnei ti timè :

! !
1 2 2 3
aij = aij =
1
2 3 2
3 4
ParathroÔme ìti isqÔei h akìloujh sqèsh: ai i i = ai i i
1 2 3 2 3 1.
Epsh parathretai ìti h para-

pˆnw sqèsh enai dunatìn na isqÔsei an gnoun dÔo antimetajèsei deikt¸n ston ai i i i ;i
1 2 3: ( 1 2)

kai ( i ;i
1 3 ). 'Ara, o A enai summetrikì anˆ dÔo.

2.3.11 Tanustè w Grammiko Metasqhmatismo

Oi tanustè uyhl  tˆxh mporoÔn na qrhsimopoihjoÔn gia metasqhmatismoÔ anˆmesa se

dianusmatikoÔ q¸rou (kai tanustikoÔ q¸rou ), gia parˆdeigma anˆmesa se dianÔsmata,

  anˆmesa se ènan pnaka kai ènan tanust  3h tˆxh . Profan¸ , apotele ma genkeush

twn grammik¸n metasqhmatism¸n me pnake , oi opooi mporoÔn na qrhsimopoihjoÔn gia na

sundèsoun dianusmatikoÔ q¸rou .

Orismì 2.3.26. Grammikì Metasqhmatismì Tanust¸n


( ) JewroÔme tou tanustè

A2 C J1 J2 Jn1 kai B2 C I1 I2 In2 . O grammikì metasqhmatismì tou A ston B


pragmatopoietai qrhsimopoi¸nta ton tanust  C 2 C I I In J J Jn
1 2 2 1 2 1:

B = hC ; Ain 2 +1 ;:::;n2 +n1 ;1;:::;n1 : (2.41)

H sqèsh (2.41) mpore na grafe enallaktikˆ:

J1 X
X J2 X
Jn
1

bi i :::in =
1 2 2
 i i :::in
1 2 j j :::jn1 aj1 j2 :::jn1 :
2 1 2
(2.42)

j1 =1 j2 =1 jn1 =1

21
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Ston parapˆnw orismì, o tanust  C mpore na qrhsimopoihje gia na anaparast sei pol-

lè diaforetikè sundèsei , anˆloga me tou dekte pou qrhsimopoioÔntai sta ajrosmata

th (2.42). Na shmeiwje ìti o parapˆnw metasqhmatismì mpore na ulopoihje qrhsimopoi¸-

nta ta anaptÔgmata twn tanust¸n. Se aut n thn perptwsh, to anˆptugma tou B mpore na

upologiste me to ginìmeno twn pinˆkwn pou ekfrˆzoun ta anaptÔgmata twn A kai C . Na sh-

meiwje epsh ìti sthn perptwsh twn mikt¸n tanust¸n, o metasqhmatismì gnetai anˆmesa

se summetablhtoÔ kai antimetablhtoÔ dekte twn A kai C .

2.4 Statistikˆ Mètra Uyhlìterh Tˆxh

H paroÔsa enìthta anafèretai ston orismì statistik¸n mètrwn uyhlìterh tˆxh ( higher-
order statisti s ). Sundèontai me to pedo th anˆlush tanust¸n, kaj¸ oi ropogenn trie

kai susswreÔtrie enai sthn ousa uper-summetriko tanustè uyhl  tˆxh . Perigrafè twn

statistik¸n uyhlìterh tˆxh upˆrqoun sta [51, 73℄.

2.4.1 Ropogenn trie kai SusswreÔtrie

Orismì 2.4.1. Apì koinoÔ Ropogenn tria 1h Tˆxh


( ) JewroÔme èna sÔnolo apì n
tuqae metablhtè x ; x ; : : : ; xn . H apì koinoÔ ropogenn tria (1h qarakthristik  sunˆrth-
1 2

sh) (! ; ! ; : : : ; !n ), ! 2 R orzetai w :


1 2

(! ; ! ; : : : ; !n) = fej ! x


1 2 E
( !x
1 1+ 2 2+  !n xn
+ )
g (2.43)

Praktikˆ, h apì koinoÔ 1h qarakthristik  sunˆrthsh mpore na jewrhje w o suzug  me-

tasqhmatismì Fourier th apì koinoÔ sunˆrthsh puknìthta pijanìthta p(x ; x ; : : : ; xn ):


1 2
Z + 1 Z + 1 Z +1
(! ; ! ; : : : ; !n) =
1 2  p(x ; x ; : : : ; xn )ej ! x
1 2
( !x
1 1+ 2 2+  !n xn dx dx : : : dx
+ )
1 n 2
1 1 1
(2.44)

Orismì 2.4.2. Apì koinoÔ Ropogenn tria 2h Tˆxh


( ) H apì koinoÔ ropogenn tria

(2h qarakthristik  sunˆrthsh) (! ; ! ; : : : ; !n) ! 2 R


1 2 , orzetai w :

(! ; ! ; : : : ; !n) = ln (! ; ! ; : : : ; !n)


1 2 1 2 (2.45)

Orismì 2.4.3. Apì koinoÔ Rop 


( ) JewroÔme èna sÔnolo apì n tuqae metablhtè
x ; x ; : : : ; xn .
1 2 H apì koinoÔ rop  ( joint moment ) tˆxh r = k + k +    + kn twn tu-
1 2

qawn metablht¸n orzetai w :



 r (!
; ! ; : : : ; !n)
mk k :::kn = Ef xk11 xk22 : : : xknn g=( j )r 1 2

! ! k    !nkn !
(2.46)
1 2 k
1
1
2
2
!
1= 2=  !n
= =0

22
2.4. STATISTIKŸ
A MŸ
ETRA UYHLŸ
OTERHS TŸ
AXHS

Orismì 2.4.4. Apì koinoÔ SusswreÔtria


( n tuqae meta-
) JewroÔme èna sÔnolo apì

blhtè x ; x ; : : : ; xn . H apì koinoÔ susswreÔtria (joint umulant) tˆxh r = k + k +    + kn


1 2 1 2

twn tuqawn metablht¸n orzetai w :


 r (!
; ! ; : : : ; !n)
k k :::kn = ( j )r 1 2

! k ! k    !nkn !
1 2
(2.47)
1
1
2
2
1= 2= !  !n
= =0

Na shmeiwje epsh ìti gia diaforetikè tuqae metablhtè h rop  deÔterh tˆxh a-

nafèretai w eterosusqètish ( ross- orrelation ), en¸ h rop  deÔterh tˆxh gia ma tuqaa

metablht  anafèretai w susqètish ( orrelation ). Gia ma tuqaa metablht  x 1, oi pr¸te 4

ropè enai oi:

m = Efx g;
1 1 m = Efx 2
2
1
g
m = Efx g;
3
3
1
m = Efx 4
4
1
g
kai sundèontai me ti antstoiqe susswreÔtriè tou me ti akìlouje sqèsei :

=m ;
1 1 =m2 2 m 2
1

= m 3m m + 2m ;
3 3 2 1
2
1
=m4 4 4m m 3 1 3m + 12m m
2
2 2
2
1
6m 4
1

Orismì 2.4.5. Rop  N -ost  Tˆxh Tuqaou DianÔsmato


( ) JewroÔme tuqao diˆ-

nusma x. H rop  N -ost  tˆxh tou x orzetai w :

MN = fx
x
  
xg
E (2.48)

Praktikˆ, h rop  N -ost  tˆxh tuqaou dianÔsmato enai tanust  N -ost  tˆxh . Gia

migadikˆ tuqaa dianÔsmata, isqÔoun oi parakˆtw sqèsei gia ti ropè tˆxh 1 èw 4:

m = Efxg;
1 M = fx
x g
2 E

M = Efx
x
xg;
3 M = fx
x
x
xg
4 E

Gia stˆsime stoqastikè diadikase x(t), enai dunatìn na oristoÔn ropè me bˆsh ti

qronikè kajuster sei :

mn ( ;  ; : : : ; n ) = Efx(t)x(t +  )    x(t + n )g
1 2 1 1 1 (2.49)

23
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

Oi sqèsei pou sundèoun ti susswreÔtrie me ti ropè th x(t) gia thn pr¸th kai deÔterh
tˆxh:

= m = Efx(t)g;
1 1 ( ) = m ( ) (m )
2 1 2 1 1
2

H m ( ) onomˆzetai sthn bibliografa w autosusqètish kai h ( ) w summetablhtìthta.


2 2

An h diadikasa èqei m = 0 (mhdenikì mèso ìro), oi susswreÔtrie 1h , 2h , kai 3h tˆxh


1

isoÔntai me ti antstoiqe ropè tou . Jètonta epsh  =  =  = 0, proÔptoun ta 1 2 3

parakˆtw gnwstˆ statistikˆ mètra:

= (0) = Efx (t)g (metablhtìthta - varian e)


2 2
2

= (0) = Efx (t)g (strèblwsh - skewness)


3 3
3

= (0) = Efx (t)g 3( ) (kÔrtwsh - kurtosis)


4 4
4
2
2

2.4.2 Idiìthte Rop¸n kai Susswreutri¸n

Idiìthta 2.4.1. Klimˆkwsh ( ) JewroÔme tuqae metablhtè x ; x ; : : : ; xn .


1 2 An oi tuqae

metablhtè pollaplasiastoÔn me stajerè ; ; : : : ; n 2 R, tìte:


1 2

n
Y
m( x ; x ; : : : ; n xn ) =
1 1 2 2 i  m(x ; x ; : : : ; xn )
1 2 (2.50)

i=1
Yn
( x ; x ; : : : ; n xn ) =
1 1 2 2 i  (x ; x ; : : : ; xn )
1 2 (2.51)

i=1

Idiìthta 2.4.2. 'Ajroisma ( ) JewroÔme tuqae metablhtè y; x ; x ; : : : ; xn .


1 2 IsqÔoun oi

parakˆtw sqèsei gia ti apì koinoÔ ropè kai susswreÔtrie gia ti tuqae metablhtè

x + y; x ; : : : ; xn :
1 2

m(x + y; x ; : : : ; xn ) = m(x ; x ; : : : ; xn ) + m(y; x ; : : : ; xn )


1 2 1 2 2 (2.52)

(x + y; x ; : : : ; xn ) = (x ; x ; : : : ; xn ) + (y; x ; : : : ; xn )
1 2 1 2 2 (2.53)

Idiìthta 2.4.3. Polugrammikìthta


( ) JewroÔme pragmatikì tuqao diˆnusma x, me stoi-
qeax ; x ; : : : ; xn . An to x metasqhmatiste sto x~ pollaplasiˆzontˆ to me pnaka A, ètsi
1 2

~ = A  x, tìte isqÔoun oi akìlouje sqèsei gia ti ropè kai susswreÔtrie tou x~ :


¸ste x

Mx = Mx  A  A    N A
~
N N 1 (2.54)
2

CNx = CNx 
~
1 A  A    N A
2 (2.55)

gia kˆje tˆxh N.

24
2.4. STATISTIKŸ
A MŸ
ETRA UYHLŸ
OTERHS TŸ
AXHS

Idiìthta 2.4.4. Summetra ( ) JewroÔme tuqae metablhtè x ; x ; : : : ; xn .


1 2 Oi apì koinoÔ

ropè kai susswreÔtrie twn x ; x ; : : : ; xn enai summetrikè :


1 2

m(x ; x ; : : : ; xn ) = m(xP ; xP ; : : : ; xP n )
1 2 (1) (2) ( ) (2.56)

(x ; x ; : : : ; xn ) = (xP ; xP ; : : : ; xP n )
1 2 (1) (2) ( ) (2.57)

ìpou P enai ma antimetˆjesh twn 1; 2; : : : ; n .

Basik  sunèpeia th Idiìthta 2.4.4 enai ìti oi ropè kai susswreÔtrie tuqawn dianu-

smˆtwn enai uper-summetriko tanustè .

Idiìthta 2.4.5. 'Artia Sunˆrthsh Puknìthta Pijanìthta


( ) JewroÔme tuqaa meta-

blht  y. An h sunˆrthsh puknìthta pijanìthta th y , p(y ) enai ˆrtia, tìte oi ropè kai
oi susswreÔtrie peritt  tˆxh th y mhdenzontai.
Idiìthta 2.4.6. Metatìpish ( ) JewroÔme tuqae metablhtè x ; x ; : : : ; xn
1 2 kai stajerˆ

2 R. IsqÔoun oi akìlouje idiìthte :

m( + x ; x ; : : : ; xn ) = m(x ; : : : ; xn ) + m(x ; : : : ; xn )
1 2 2 1 (2.58)

( + x ) = + (x )
1 1 (2.59)

( + x ; x ; : : : ; xn ) = (x ; x ; : : : ; xn )
1 2 1 2 (2.60)

Idiìthta 2.4.7. Anexˆrthte Tuqae Metablhtè


( ) JewroÔme tuqae metablhtè x;
1

x ; : : : ; xn .
2 An èna uposÔnolo twn n tuqawn metablht¸n enai anexˆrthto w pro ti upì-
loipe , tìte h apì koinoÔ susswreÔtria isoÔtai me mhdèn.

San sunèpeia th idiìthta 2.4.7, h susswreÔtria uyhl  tˆxh enì tuqaou dianÔsmato

me amoibaa anexˆrthta stoiqea enai diag¸nio tanust  .

Idiìthta 2.4.8. 'Ajroisma Anexˆrthtwn Tuqawn Metablht¸n


( ) JewroÔme tuqa-

e metablhtè x ; x ; : : : ; xn ; y ; y ; : : : ; yn.
1 2 1 2 An oi x ; x ; : : : ; xn
1 2 enai anexˆrthte apì ti

y ; y ; : : : ; yn, tìte isqÔei h parakˆtw sqèsh:


1 2

(x + y ; x + y ; : : : ; xn + yn ) = (x ; x ; : : : ; xn ) + (y ; y ; : : : ; yn)
1 1 2 2 1 2 1 2 (2.61)

Idiìthta 2.4.9. Mh Kanonikìthta


( x kai tuqaa metablht 
) JewroÔme tuqaa metablht 

y pou akolouje thn kanonik  katanom . H mèsh tim  kai diakÔmansh th y isoÔntai me autè
th x. IsqÔei h akìloujh sqèsh gia tˆxh N > 3:

xN = mxN myN (2.62)

25
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

San sunèpeia th parapˆnw idiìthta , oi susswreÔtrie uyhl  tˆxh tuqaa metablht 

pou akolouje kanonik  katanom  isoÔntai me 0. Genikˆ, oi susswreÔtrie tˆxh N >2 den

ephreˆzontai apì th mèsh tim  tuqawn metablht¸n. Epsh , oi idiìthte twn susswreutri¸n

gia kanonikè tuqae metablhtè , kaj¸ kai gia amoibaa anexˆrthte metablhtè , odhgoÔn

sto sumpèrasma ìti enai protimìterh h qr sh susswreutri¸n parˆ rop¸n.

2.4.3 Upologismì Statistik¸n Mètrwn Uyhlìterh Tˆxh

JewroÔme thn tuqaa metablht  x, kai T anexˆrthta degmata th x, ta opoa sumbolzontai


me xt (1 6 t 6 T ). O upologismì th rop  N -ost  tˆxh mxN dnetai apì ton tÔpo:

T
^ xN =
m
1X xN
T t (2.63)

t=1

ApodeiknÔetai ìti gia T !1 , h ektmhsh th rop  ^ xN


m sugklnei sthn pragmatik  rop 

mxN , me pijanìthta 1. Epsh isqÔei ìti E fm^ xN g = mxN . H diakÔmansh th ^ xN


m dnetai apì

ton tÔpo:

Varfm^ xN g = T1 [mxN (mXN ) ℄


2
2
(2.64)

Tèlo , apodeiknÔetai ìti h ^ xN


m asumptwtikˆ akolouje thn kanonik  katanom . Parathr -

sei sqetikˆ me thn ektmhsh rop¸n kai idiìthte twn ektim¸menwn rop¸n uyhlìterh tˆxh

paratjentai sto [14℄.

Gia ton upologismì twn susswreutri¸n, èqei protaje h qr sh twn k-statistik¸n, pou
apoteloÔn mh polwmènou ektimhtè susswreutri¸n. Gia ton orismì twn k -statistik¸n, qrh-

simopoioÔntai oi ektim¸mene kentrikè ropè :

T
~ xN =
m
1X (x ^ x)N
m
t (2.65)
T t=1
1

Ta k-statistikˆ pou ektimoÔn susswreÔtrie tˆxh 1 èw 4 enai ta akìlouja:

kx = ^x = m
1
^x
1 1

T
kx = ^x = ~
m x
2
T 1
2 2

T 2

kx = ^x =
2) m~
x
3
(T 1)(T
3 3

 
T 2

kx4
= ^x
4
= (T 1)(T 2)(T 3) (T + 1)m~ x 4
3(T 1)(m~ x)2
2

26
2.5. APOKLŸ
ISEIS BREGMAN

ApodeiknÔetai ìti gia T ! 1 ^xN , h sugklnei sthn xN me pijanìthta 1. Epsh isqÔei ìti

E f g=
^xn xN . Oi diakumˆnsei th ^xN gia tˆxei 1 èw 4 dnontai parakˆtw:

x
Varf ^xg = T
1
2

Varf ^xg = T + 2( x)
x 2
4 2
2
T 1
Varf ^xg = T + T + 9(1 ) + (T 6T1)(
9 ( x)
x x x x 2 3
6 4 2 3 2
3
T 2)
Varf ^xg = T + 16 + 48T 1 + 34( )
x x x x x x 2
8 6 2 5 3 4
4

+ 72T (T( )1)(+T144(2) ) + (T 241)(


T (T + 1)( x)
x x x x2 2 4
4 2 3 2 2

T 2)(T 3)

Tèlo , apodeiknÔetai ìti h ^xN asumptwtikˆ akolouje thn kanonik  katanom . H diakÔmansh

th ektim¸menh susswreÔtria uyhl  tˆxh epitrèpei thn ektmhsh tou arijmoÔ deigmˆtwn

pou apaitetai gia ton upologismì susswreÔtria me dedomènh akrbeia.

2.5 Apoklsei Bregman

H paroÔsa enìthta anafèretai se ma oikogèneia sunart sewn pou enai gnwstè w apokl-

sei Bregman . Protˆjhkan apì ton Bregman to 1967 [16℄. Ja dojoÔn oi basiko orismo

kai idiìthte , kaj¸ kai oi sqèsei pou anafèrontai sqetikˆ me ti apoklsei Bregman ston

tanustikì q¸ro. Oi sqèsei pou anafèrontai ja qrhsimopoihjoÔn gia thn eÔresh tou algo-

rjmou paragontopohsh mh arnhtik¸n tanust¸n, pou ja perigrafe sto Kefˆlaio 4.

Orismì 2.5.1. Apìklish Bregman


(  : S  R ! R , me
) JewroÔme thn kurt  sunˆrthsh

H apìklish Bregman th  enai h sunˆrthsh D : S  int(S ) ! R ,


+
suneq  pr¸th parˆgwgo.

kai orzetai w :

D (x; y ) = (x) (y ) 0 (y )(x y ) (2.66)

ìpou S sÔnolo, int(S ) to eswterikì sÔnolo (interior set) tou S , kai 0 (y ) h pr¸th parˆgwgo
th .

Oi apoklsei Bregman parousiˆzoun endiafèrouse idiìthte [83℄. Enai èqoun mh ar-

nhtikè timè , o pr¸to tou ìro enai kurtì ( onvex x=y


), kai mhdenzontai mìno ìtan .

Epsh , isqÔei h idiìthta th grammikìthta : D  (x; y ) = D (x; y ) + D (x; y )


1+ 2 1 2
. 'Ara, to

ˆjroisma dÔo apoklsewn Bregman enai epsh apìklish Bregman . AkoloÔjw , o orismì

th apìklish Bregman mpore na epektaje gia stoiqea pinˆkwn:

X
D (X; Y) = D (xij ; yij ) (2.67)

ij

27
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

ìpou X; Y pnake . Antstoiqa, an jewr soume tanustè X ; Y 2 R I I IN


1 2
, o orismì

mpore na epektaje akoloÔjw :

2 IN
I1 IX
D (X ; Y ) = D (xi i :::iN ; yi i :::iN )
1 2 1 2
(2.68)

i1 i2 :::iN

Paradegmata tupik¸n (x) pou qrhsimopoioÔntai gia thn dhmiourga apoklsewn Bergman
paratjentai ston Pnaka 2.1 [4℄ [83℄. H Nìrma Frobenius 2ou BajmoÔ,   alli¸ tetragwnik 

Eukledia apìstash, enai h pio eurèw qrhsimopoioÔmenh apìklish Bregman . Enai kurt 

kai ston 2o ìro th . Sto sq ma 2.1 apeikonzetai h diadikasa upologismoÔ th apìklsh

Bregman gia thn nìrma Frobenius .

H apìklish Kullba k-Leibler qrhsimopoietai sthn jewra pijanot twn gia na entopiste

h diaforˆ metaxÔ dÔo katanom¸n puknìthta pijanìthta . Kai h apìklish Kullba k-Leibler
enai kurt  ston 2o ìro th . H apìstash Itakura-Saito qrhsimopoietai sto pedo th epexer-

gasa shmˆtwn gia sÔgkrish fasmˆtwn isqÔo .

Pnaka 2.1: Tupikè apoklsei Bregman .

'Onoma Apìklish Pedo (x) D (x; y )


Nìrma Frobenius 2ou BajmoÔ R x 1
2
2
(x y) 1
2
2

Apìklish Kullba k-Leibler R+ x log(x) x log( xy ) x + y


Apìstash Itakura-Saito R+ log(x) x log( x ) 1
y y
Logistik  Ap¸leia ( logisti loss
) [0; 1℄ x log(x) + (1 x) log(1 x) x log( xy ) + (1 x) log( 1
1
x)
y
Ekjetik  R ex ex ey (x y )ey

Sq ma 2.1: Apeikìnish upologismoÔ th apìklish Bregman gia (z ) = z1


2
2
(apì to [83℄).

28
2.5. APOKLŸ
ISEIS BREGMAN

Sq ma 2.2: Apeikìnish upologismoÔ th apìklish Bregman gia (z ) = z log(z ) (apì to

[83℄).

29
KEFŸ
ALAIO 2. LOGISMŸ
OS TANUSTŸ
WN

30
Kefˆlaio 3

Anagn¸rish MousikoÔ Edou

3.1 Eisagwg 

Ta teleutaa qrìnia parathretai h dhmiourga megˆlwn bˆsewn dedomènwn, proerqìmene

tìso apì thn anˆkthsh analogik¸n arqewn ìso kai apì nèa yhfiakˆ dedomèna. Kajstatai

loipìn epitaktik  h anˆgkh gia th dhmiourga axiìpistwn kai gr gorwn ergalewn gia anˆ-

lush dedomènwn gia na qrhsimopoihje se anaz thsh kai prìsbash se ulikì. Sto pedo th

anaz thsh kai anˆkthsh dedomènwn, ta mousikˆ edh ( musi al genres) èqoun megˆlo rìlo,

kaj¸ qrhsimopoioÔntai ed¸ kai pollˆ qrìnia sthn orgˆnwsh mousik¸n katalìgwn, mousik¸n

katasthmˆtwn, kai se yhfikakè biblioj ke polumesik¸n dedomènwn [78℄. Qarakthristikì

enai ìti oi katˆlogoi mousik¸n kommati¸n megal¸noun me megˆlou rujmoÔ , xepern¸nta

katˆ polÔ to 1 ekatommÔrio kommˆtia. Me bˆsh ta parapˆnw stoiqea, h susqètish enì

mousikoÔ kommatioÔ ( musi al tra k ) me kˆpoio mousikì edo enai shmantik  sto na bohj sei

tou qr ste -pelˆte sthn anaz ths  tou .

Parìla autˆ, to prìblhma th anagn¸rish mousik¸n eid¸n den enai tetrimmèno. En¸

upˆrqei diadedomènh qr sh twn mousik¸n eid¸n, autˆ paramènoun ma kak¸ orismènh ènnoia.

Enai qarakthristikì ìti h swst  taxinìmhsh mousikoÔ edou apì anjr¸pou (mh eidikoÔ

sthn taxinìmhsh mousik¸n eid¸n) ftˆnei polÔ qamhlˆ posostˆ: gia to prìblhma th taxinì-

mhsh kommati¸n se 10 mousikˆ edh, h anjr¸pinh akrbeia enai mìli 53% gia degmata 250

mse , kai anebanei sto 72% gia degmata 3 se [72℄. Tautìqrona, en¸ polÔ suqnˆ qrhsimo-

poioÔntai ìroi, ìpw pop ro k


, , kai jazz
, auto enai asaf¸ orismènoi kai pollè forè èna

kommˆti antistoiqe se perissìtera apì èna mousikˆ edh. Tèlo , tjetai to z thma poÔ ja

efarmoste h taxinìmhsh mousikoÔ edou : sto kommˆti, sto ˆlmpoum (sullog  kommati¸n)

  ston kallitèqnh. Sta perissìtera hlektronikˆ katast mata h taxinìmhsh gnetai me bˆsh

to ˆlmpoum. 'Omw , gnetai katanohtì ìti afoÔ èna kommˆti mpore na an kei se perissìtera

apì èna edh, tìte èna ˆlmpoum, to opoo mpore kai na perièqei eterogenè ulikì, den enai

31
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

dunatìn na an kei se èna mìno edo .

'Ara, gia to prìblhma th taxinìmhsh mousik¸n eid¸n, apaitetai ma ierarqik  taxinìmhsh

twn mousik¸n eid¸n, h opoa na kalÔptei ìso to dunatìn gnetai ìla ta edh, kai parˆllhla

na mhn upˆrqoun pollè epikalÔyei kai asˆfeie stou orismoÔ twn eid¸n. O Pa het ìmw

èdeixe ìti h dhmiourga ma tètoia ierarqa den enai aplì z thma [64℄. Gia parˆdeigma, to

site Amazon1 èqei kathgoriopoi sei kommˆtia qrhsimopoi¸nta 719 mousikˆ edh. To site All-
musi 2 qrhsimopoie 531 edh kai to Mp3 3 qrhsimopoie 430 edh. Mìno 70 mousikˆ edh  tan

koinˆ kai sta tra sites . Epsh , parathretai apì ton Pa het ìti oi asˆfeie kai diaforetikè

taxinom sei eid¸n den apoteloÔn kat anˆgkh prìblhma gia tou qr ste , allˆ profan¸

den prosfèrontai gia qr sh se sust mata autìmath taxinìmhsh [64℄. 'Ena ˆllo prìblhma

prostjetai sto prìblhma eÔresh taxinìmhsh : ìti exartˆtai apì thn perioq  sthn opoa

protenetai. Gia parˆdeigma, èna kommˆti ènteqh ellhnik  mousik  se mh ellhnikè uphre-

se ja taxinomhje w mousik  tou kìsmou ( world musi ). Epsh , upˆrqei asˆfeia me bˆsh

ta krit ria apì ta opoa prokÔptei ma ierarqa: ˆlle ierarqe prokÔptoun qarakthrzonta

qronikè periìdou (pq. tragoÔdia dekaeta tou 50), ˆlle me bˆsh th q¸ra, me bˆsh th

gl¸ssa, me bˆsh to jèma twn kommati¸n,   me bˆsh ton kallitèqnh. 'Ena teleutao prìblhma

pou prokÔptei, enai h prosj kh nèou mousikoÔ edou sthn  dh upˆrqousa ierarqa. Ta nèa

mousikˆ edh prokÔptoun sun jw sugqwneÔonta  dh upˆrqonta edh,   qwrzonta èna  dh

upˆrqon edo se kathgore . To parapˆnw jèma apotele prìblhma gia èna sÔsthma autì-

math taxinìmhsh mousik¸n eid¸n, to opoo ja prèpei na metabˆllei dunamikˆ thn ierarqa

tou.

To parìn kefˆlaio suneqzei me ma anaskìphsh twn dedomènwn kai ierarqi¸n pou qrhsi-

mopoi jhkan se efarmogè autìmath taxinìmhsh mousik¸n eid¸n sthn enìthta 3.2. Sthn

enìthta 3.3 perigrˆfontai ta qarakthristikˆ pou exˆgontai apì thn kumatomorf  tou mousi-

koÔ kommatioÔ gia thn perigraf  mousikoÔ edou . Sthn enìthta 3.4 perigrˆfontai sunoptikˆ

algìrijmoi taxinìmhsh pou èqoun qrhsimopoihje gia thn eplush tou probl mato th au-

tìmath taxinìmhsh mousikoÔ edou . Tèlo , h enìthta 3.5 parajètei telikˆ sumperˆsmata

gia to pedo.

3.2 SÔnola dedomènwn kai Ierarqe

O Pa het apopeirˆjhke to 2000 na dhmiourg sei ma exantlhtik  ierarqa mousik¸n eid¸n,

prospaj¸nta parˆllhla na mhn upˆrqoun allepikalÔyei se mousikˆ edh, na dnetai h du-

natìthta prosj kh nèou mousikoÔ edou , kai h ierarqa na enai antikeimenik , dhlad  na

1 http://www.amazon. om
2 http://www.allmusi . om
3 http://www.mp3. om

32
3.2. SŸ
UNOLA DEDOMŸ
ENWN KAI IERARQŸ
IES

mhn ephreˆzetai apì qronikoÔ kai topikoÔ parˆgonte [64℄. H ierarqa pou dhmiourg jhke

eqe san pr¸to eppedo basikˆ mousikˆ edh kai se deÔtero eppedo perigrafè exeidkeush

tou mousikoÔ edou (oi perigrafè anafèrontai se q¸ra, enorq strwsh, kai kallitèqnh).

Sunolikˆ h ierarqa pou protˆjhke perieqe 378 mousikˆ edh. 'Omw , telikˆ o Pa het e-

gkatèleiye th dhmiourga th ierarqa [3℄, lìgw poll¸n epikalÔyewn pou up rqan anˆmesa

se edh kai se euaisjhsa th taxinìmhsh se nèa mousikˆ edh pou prokÔptoun apì mexh

eid¸n pou up rqan sthn ierarqa. H nèa ierarqa pou protˆjhke apì ton Pa het - gia to

prìgramma anaz thsh mousik¸n eid¸n tou ereunhtikoÔ èrgou Cuidado - eqe 2 eppeda, to

pr¸to me 18 genikˆ mousikˆ edh kai to deÔtero me 250 upo-edh [65℄. Ta 18 genikˆ mousikˆ

Ambien e, Blues, Classi al, Country, Ele troni a, Folk, Hard, Hip Hop, Jazz, New
edh enai:

Age, Pop, Reggae, Rhythm&Blues, Ro k, Ro k&Roll, Soul, Variety, World kai .

Oi Tzanetˆkh kai Cook dhmioÔrghsan to 2002 ma bˆsh dÔo epipèdwn gia peirˆmata

anagn¸rish mousikoÔ edou (GTZAN dataset ) [84℄. Sto pr¸to eppedo upˆrqoun 10 basikˆ

mousikˆ edh, en¸ to deÔtero eppedo perièqei 4 upo-edh gia thn kathgora Classi al kai 6

upo-edh gia thn kathgoraJazz . Gia kˆje mousikì edo kai upo-edo dhmiourg jhkan 100

se
arqea, diˆrkeia 30 to kajèna. Na shmeiwje ìti ta peirˆmata genik  taxinìmhsh pou

pragmatopohse o Tzanetˆkh sto [84℄ èginan xeqwristˆ apì ta peirˆmata taxinìmhsh sta

upo-edh th klassik  kai jazz mousik  . H ierarqa th bˆsh dedomènwn tou Tzanetˆkh

apeikonzetai sto Sq ma 3.1. Peirˆmata sthn bˆsh tou Tzanetˆkh pragmatopoi jhkan apì

ton dio sto [84℄, apì ton Li sto [57℄, apì ton Lidy sto [58℄, kai pragmatopoioÔntai kai sthn

paroÔsa diatrib .

Choir
Classi al Or hestra
Country Piano
Dis o String Quartet
HipHop Bigband
Jazz Cool
Ro k Fusion
Blues Piano
Reggae Quartet
Pop Swing
Metal

Sq ma 3.1: Ierarqa taxinìmhsh mousik¸n eid¸n tou Tzanetˆkh [84℄.

O Ogihara sta peirˆmata pou pragmatopohse gia anagn¸rish mousik¸n eid¸n, dhmioÔr-

ghse ma bˆsh qrhsimopoi¸nta mousikˆ kommˆtia apì proswpikè sullogè [57℄. H bˆsh

33
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

perieqe 756 arqea  qou, ta opoa sugkentr¸jhkan apì 189 mousikˆ ˆlmpoum. Ta arqea

 tan organwmèna se 5 mousikˆ edh: Ambient, Classi al, Fusion, Jazz, Ro k kai . Kˆje arqeo

eqe diˆrkeia 30 se
. 109 arqea dhmiourg jhkan gia thn klˆsh Ambient , 164 gia thn Classi-
al , 136 gia thn Fusion , 251 gia thn Jazz
, kai 96 gia thn Ro k . Ta arqea èqoun suqnìthta

deigmatolhya 22.050 Hz , sta 16 bit


, monofwnikˆ. Na shmeiwje ìti h bˆsh dedomènwn pou

anaptÔqjhke apì ton Ogihara den perilambˆnei arketˆ mousikˆ edh gia na qarakthriste

axiìpisth. Epsh , upˆrqei asˆfeia ston orismì twn eid¸n: to fusion apoteletai apì th

sugq¸neush dÔo   perissìterwn mousik¸n eid¸n kai den prèpei na oriste apì mìno tou san

edo . Antstoiqa, to Ambient perigrˆfei perissìtero to stul th mousik  parˆ to edo

th (kommˆtia pou enai sthn kathgora Classi ial kai Jazz ja mporoÔsan na qarakthristoÔn

w Ambient ).

Gia to sunèdrio ISMIR 2004


4
dhmiourg jhkan 2 sÔnola dedomènwn, gia qr sh se diagw-

nismì anagn¸rish mousikoÔ edou . To pr¸to sÔnolo onomˆzetai ISMIRgenre kai perièqei

1458 arqea me genikè klˆsei mousik¸n eid¸n, kalÔptonta 6 genikè klˆsei . To deÔtero

sÔnolo perieqe mousikˆ edh qor¸n, gia taxinìmhsh kommati¸n me bˆsh ton rujmì. Onomˆze-

tai ISMIRrhythm kai perièqei 698 arqea pou kalÔptoun 8 qoreutikˆ mousikˆ edh. Ta edh

pou kalÔptei h bˆsh ISMIRgenre parousiˆzontai ston Pnaka 3.1 kai ta edh pou kalÔptei

h bˆsh ISMIRrhythm ston Pnaka 3.2. Peirˆmata kai sthn bˆsh ISMIRgenre pragmatopoi-

 jhkan sta plasia tou diagwnismoÔ ISMIR 2004 apì tou Lidy Ellis West Pampalk
, , , , kai

Tzanetˆkh [58℄. Peirˆmata sthn bˆsh ISMIRrhythm pragmatopoi jhkan apì tou Dixon
[25℄, Gouyon [36℄, kai Lidy [58℄.

Antstoiqh bˆsh dhmiourg jhke gia to sunèdrio ISMIR 5


2005 , ìpou dhmiourg jhkan dÔo

bˆsei dedomènwn apì diaforetikè phgè . H pr¸th bˆsh apoteletai apì 1.515 kommˆtia pou

kalÔptoun 10 mousikˆ edh: lassi al, ambient, ele troni , new-age, ro k, punk, jazz, blues,
folk, kaiethni . H deÔterh bˆsh apoteletai apì 1.414 kommˆtia pou kalÔptoun 6 mousikˆ

edh: ro k, hip-hop, ountry, new-age, kai reggae . Sugkentrwtikˆ apotelèsmata gia ta dÔo

sÔnola dedomènwn tou diagwnismoÔ upˆrqoun sto [78℄.

To 2005 o Meng dhmioÔrghse dÔo sÔnola dedomènwn gia peirˆmata se anagn¸rish mou-

sikoÔ edou [62℄. To pr¸to sÔnolo perieqe 100 kommˆtia, organwmèno se 5 edh: lassi al,
hard ro k, jazz, pop , kai te hno . To deÔtero sÔnolo dedomènwn apoteletai apì 354 degmata

diˆrkeia 30 se to kˆje èna, proerqìmeno apì dwreˆn degmata pou parèqei to Amazon . To

deÔtero sÔnolo organ¸netai se 6 mousikˆ edh: lassi al, ountry, jazz, rap, ro k te hno , kai .

O Burred prìteine th dhmiourga ma ierarqik  orgˆnwsh mousik¸n eid¸n me 3 eppeda

[19℄. Ta pr¸ta 2 eppeda apeikonzontai sto Sq ma 3.2. To edo Chamber Musi qwrzetai

4 http://ismir2004.ismir.net/ISMIR Contest.html
5 http://www.musi -ir.org/evaluation/mirex-results/audio-genre/index.html

34
3.2. SŸ
UNOLA DEDOMŸ
ENWN KAI IERARQŸ
IES

Mousikì Edo Arijmì Arqewn


Classi al 640

Ele troni 229

Jazz - Blues 52

Metal - Punk 90

Ro k - Pop 203

World 244

Pnaka 3.1: Ta mousikˆ edh pou kalÔptontai apì th bˆsh ISMIRgenre .

Mousikì Edo Arijmì Arqewn


Cha Cha Cha 111

Jive 60

Qui kstep 82

Rumba 98

Samba 86

Slow Waltz 110

Tango 86

Viennese Waltz 65

Pnaka 3.2: Ta mousikˆ edh pou kalÔptontai apì th bˆsh ISMIRrhythm .

sta: Chamber musi with piano, Solo musi , String quartet, Other hamber ensembles .

To edo Or hestral musi Symphoni musi , Or hestra with hoir,


qwrzetai sta upo-edh:

Or hestra with soloist . Ro k


To edo Hard ro k
qwrzetai sta: Soft ro kkai . Tèlo ,

to edo Ele troni /Pop Te hno/Dan e, Rap/Hip-Hop, Pop


qwrzetai sta edh: . To edo

Jazz/Blues den qwrzetai se upo-edh. Burred


Gia ta 13 upo-edh pou protˆjhkan apì ton

dhmiourg jhke ma bˆsh dedomènwn me 50 mousikˆ kommˆtia se kˆje kathgora, ìpou to kˆje

kommˆti eqe diˆrkeia 30 se .

Musi

Classi al Non- lassi al


Chamber musi Or hestral Musi Ro k Ele troni /Pop Jazz/Blues

Sq ma 3.2: Ierarqa taxinìmhsh mousik¸n eid¸n tou Burred [19℄.

To 2007 o Barbedo prìteine ma ierarqik  orgˆnwsh mousik¸n eid¸n me 4 eppeda [5℄.

35
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

To pr¸to eppedo qwrzei th mousik  se polÔ genikˆ edh: Classi al, Pop/Ro k, Dan e . To

deÔtero eppedo exeidikeÔei: thn Classi al Instrumental, Vo al


se Pop/Ro k Organi ,
, thn se

Ele troni , kai thn Dan e se Vo al, Per ussion . Sto 4o eppedo, katal gei se 29 mousikˆ

edh.

3.3 Qarakthristikˆ Perigraf  MousikoÔ Edou

H enìthta anafèretai se qarakthristikˆ pou epitrèpoun thn perigraf  tou mousikoÔ edou .

Se orismène peript¸sei mousik  , sugkekrimèna sthn orqhstrik  dutik  mousik , enai du-

natìn h mousik  plhrofora na perigrafe qrhsimopoi¸nta anaparastˆsei uyhloÔ epipèdou,

ìpw h Musi XML6   tÔpou dedomènwn ìpw ta arqea MIDI kai .sib (tou progrˆmmato

sÔnjesh Sibelius ). Sthn prˆxh ìmw , den upˆrqoun anaparastˆsei uyhloÔ epipèdou   domè

perigraf  pou na kalÔptoun ìla (  ta perissìtera) mousikˆ edh, anˆ ton qrìno kai ton

tìpo.

'Ara, opoiad pote efarmog  autìmath anagn¸rish mousikoÔ edou apaite thn qr sh

th hqhtik  kumatomorf  , h opoa w yhfiak , enai se morf  deigmˆtwn. 'Omw , ta hqhtikˆ

degmata den mporoÔn na qrhsimopoihjoÔn apeujea se efarmogè anagn¸rish , lìgw tou me-

gˆlou ìgkou plhrofora pou perièqoun. Parˆllhla, h plhrofora pou upˆrqei sta degmata

th kumatomorf  enai qamhloÔ epipèdou kai den mpore na qrhsimopoihje gia shmasiologik 

anaparˆstash enì mousikoÔ edou . San apotèlesma, to pr¸to b ma sthn anagn¸rish mou-

sikoÔ edou enai h exagwg  qarakthristik¸n mesaou kai uyhloÔ epipèdou apì ta hqhtikˆ

degmata.

O Pa het qwrzei ta qarakthristikˆ pou qrhsimopoioÔntai sthn perigraf  mousikoÔ edou

se 3 kathgore [65℄:

 Qarakthristikˆ qroiˆ ( timbre )

 Qronikˆ-rujmikˆ qarakthristikˆ

 Qarakthristikˆ basismèna sth jemeli¸dh suqnìthta ( pit h ).

Antstoiqa, o S aringella pragmatopoie ma parìmoia orgˆnwsh twn qarakthristik¸n,

qwrzontˆ ta se 3 kathgore [78℄:

 Qarakthristikˆ qroiˆ ( timbre )

 Rujmikˆ qarakthristikˆ

 Melwdikˆ-armonikˆ qarakthristikˆ.

6 http://www.musi xml.org/xml.html

36
3.3. QARAKTHRISTIKŸ
A PERIGRAFŸ
HS MOUSIKOŸ
U EŸ
IDOUS

Sthn paroÔsa diatrib  ja qrhsimopoihje èna sunduasmì twn dÔo parapˆnw taxinom -

sewn qarakthristik¸n.

3.3.1 Qarakthristikˆ qroiˆ

H qroiˆ orzetai sth bibliografa w to qarakthristikì pou kˆnei dÔo  qou sthn dia jeme-

li¸dh suqnìthta kai me thn dia èntash na akoÔgontai diaforetiko [78℄. Ta qarakthristikˆ

pou perigrˆfoun thn qroiˆ exetˆzoun sun jw thn fasmatik  katanom  tou s mato , an kai

merikˆ upologzontai sto pedo tou qrìnou. O Peeters dhmioÔrghse ma exantlhtik  lsta

perigraf  qarakthristik¸n qroiˆ sto [70℄.

Qronikˆ qarakthristikˆ
Upologzontai gia kˆje qronikì plasio tou s mato :

 Rujmì mhdenism¸n ( Zero-Crossing Rate ): Upologzei ton arijmì for¸n pou mhdenze-

tai to s ma se èna qronikì parˆjuro. 'Ena enjìrubo s ma èqei megˆlh tim  rujmoÔ

mhdenism¸n. Qrhsimopoi jhke apì tou Tzanetˆkh [84℄, Meng [62℄, Burred [19℄, Li [57℄,

kai Cataltepe [20℄.

 Suntelestè grammik  prìbleyh ( Linear Predi tion CoeÆ ients ): QrhsimopoioÔntai

kurw sthn epexergasa omila . Oi suntelestè anafèrontai se èna fltro ìlo-pìlwn

pou jewretai ìti parˆgei to hqhtikì s ma, èqonta san esodo ma periodik  diègersh,

pou perièqei plhrofora gia thn jemeli¸dh suqnìthta. Qrhsimopoi jhkan apì tou

Tzanetˆkh [84℄, Pa het [3℄, kai Meng [62℄.

Qarakthristikˆ enèrgeia
Anafèrontai sto energeiakì perieqìmeno tou s mato :

 Tetragwnik  rza th enèrgeia ( Root Mean Square Energy ): Enai mètro th isqÔo

enì diast mato . Qrhsimopoi jhke apì ton Burred [19℄.

 Lìgo qamhl  enèrgeia ( Low Energy Rate ): Posostì twn plaiswn enì s mato

pou èqoun tetragwnik  rza th enèrgeia mikrìterh apì thn mèsh enèrgeia. Qrhsi-

mopoi jhke apì tou Tzanetˆkh [84℄, Meng [62℄, Burred [19℄, Li [57℄, kai Cataltepe
[20℄.

Fasmatikˆ qarakthristikˆ
Perigrˆfoun to sq ma tou fˆsmato isqÔo enì qronikoÔ plaisou:

37
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

 Kèntro bˆrou tou fˆsmato ( Spe tral Centroid ): orzetai w to kèntro bˆrou tou

fˆsmato isqÔo kai enai mètro tou fasmatikoÔ sq mato . Qrhsimopoi jhke apì tou

Tzanetˆkh [84℄, Burred [19℄, Pa het Li [3℄, Cataltepe


[57℄, kai [20℄. Parallagè tou

apoteloÔn h diasporˆ tou fˆsmato ( Spe tral Spread ) kai h epipedikìthta tou fˆsmato

(Spe tral Flatness ), pou qrhsimopoi jhkan apì ton Burred [19℄.

 Suqnìthta apìsbesh ( Spe tral Roll-o Frequen y ): Metrˆei se pìso uyhlè suqnì-

thte (sun jw 85%) sunantˆtai èna sugkekrimèno posostì th enèrgeia tou s mato .

Qrhsimopoi jhke apì tou Pa het [3℄, Tzanetˆkh [84℄, Li [57℄, Barbedo [5℄, kai Catal-
tepe [20℄.

 DiakÔmansh fˆsmato ( Spe trum Flux ): orzetai w h mèsh diakÔmansh tou fˆsmato

metaxÔ dÔo geitonik¸n diasthmˆtwn. Qrhsimopoi jhke apì tou Tzanetˆkh [84℄, Burred
[19℄, Pa het [3℄, Li [57℄, Barbedo [5℄, kai Cataltepe [20℄.

 Mel -qasmatiko suntelestè ( Mel Frequen y Cepstral CoeÆ ients ): Upologzontai apì

tou suntelestè braquqrìniou diakritoÔ metasqhmatismoÔ Fourier , oi opooi filtrˆ-

rontai apì ma trˆpeza fltrwn. Qrhsimopoi jhkan apì tou Tzanetˆkh [84℄, Burred
[19℄, Meng [62℄, Pa het [3℄, Li [57℄, kai Cataltepe [20℄.

Antilambanìmena qarakthristikˆ
Ta antilambanìmena qarakthristikˆ ( per eptual features ) upologzontai qrhsimopoi¸nta

èna montèlo th anjr¸pinh ako  :

 Antilambanìmenh èntash ( Per eptual Loudness ): qrhsimopoie thn èntash enì diast -

mato  qou metasqhmatismènh me bˆsh èna montèlo ako  . Qrhsimopoi jhke apì ton

Burred [19℄ kai ton Barbedo [5℄.

3.3.2 Rujmikˆ-qronikˆ qarakthristikˆ

O Pa het uposthrzei ìti èna sÔsthma autìmath anagn¸rish mousikoÔ edou den prèpei

na qrhsimopoie mìno qarakthristikˆ qroiˆ , allˆ na qrhsimopoie kai rujmik  plhrofora

[3℄. Akrib  orismì tou rujmoÔ den upˆrqei, allˆ anafèretai sthn ènnoia th qronik 

epanˆlhyh [78℄. Diaisjhtikˆ, enai profanè ìti to rujmikì perieqìmeno enì kommatioÔ

mpore na qrhsimopoihje gia diaqwrismì èntona rujmik¸n kommati¸n (pq. ro k ) kai kommati¸n

qwr èntonh asjhsh rujmoÔ (pq. klassikh mousik ).

O Gouyon exètase ta proteinìmena rujmikˆ qarakthristikˆ sthn bibliografa [36℄. Upˆr-

qoun pollè diaforetikè ìyei sthn anaz thsh rujmoÔ: exagwg  tempo , anaz thsh kroÔsh

38
3.3. QARAKTHRISTIKŸ
A PERIGRAFŸ
HS MOUSIKOŸ
U EŸ
IDOUS

( beat), anagwg  mètrou. Sun jw oi mèjodoi exagwg  rujmik¸n qarakthristik¸n anazh-

toÔn periodikìthte sto s ma, qrhsimopoi¸nta sun jw sunart sei autosusqètish   ton

metasqhmatismì Fourier tou s mato .

O Tzanetˆkh protenei th qr sh enì istogrˆmmato kroÔsh ( beat histogram ), to opoo

dhmiourgetai apì thn sunˆrthsh autosusqètish tou s mato [84℄. Parathr¸nta ta bˆrh

sti diˆfore periodikìthte tou s mato , exˆgetai ma rujmik  posìthta, pou suntele ston

diaqwrismì mousik¸n kommati¸n me èntono rujmì se sqèsh me autˆ pou den èqoun dunat  thn

asjhsh tou rujmoÔ. 'Ena parìmoio istìgramma qrhsimopoietai apì ton Burred , apì to opoo

exˆgei dÔo rujmikˆ qarakthristikˆ: dÔnamh kroÔsh ( beat strength ) kai rujmik  kanonikì-

thta ( rhythmi regularity ) [19℄. O Li qrhsimopoie epsh to istìgramma kroÔsh [57℄. To

upologzei qrhsimopoi¸nta thn sunˆrthsh autosusqètish sthn peribˆllousa tou s mato .

Ta rujmikˆ qarakthristikˆ pou exˆgontai enai: sqetikì plˆto twn dÔo koruf¸n tou isto-

grˆmmato , lìgo tou plˆtou twn dÔo koruf¸n, perodoi twn dÔo koruf¸n, kai sunolikì

ˆjroisma tou istogrˆmmato . O Meng , ektì apì to istìgramma kroÔsh , protenei kai th

qr sh fˆsmato kroÔsh ( beat spe trum ) to opoo qrhsimopoie tou Mel qasmatikoÔ sunte-

lestè gia ton upologismì tou [62℄. O Lidy protenei ta qarakthristikˆ rujmik¸n protÔpwn

rhythm pattern features


( ) [58℄. Ta qarakthristikˆ prokÔptoun efarmìzonta metasqhmati-

smìFourier kai diaforikì fltro stou suntelestè sugkekrimènh antilambanìmenh ènta-

spe i loudness sensation oeÆ ients


sh ( ). Epsh qrhsimopoietai to istìgramma rujmoÔ

rhythm histogram
( ), to opoo perièqei plhrofora rujmoÔ anˆ suqnotik  mpˆnta. Tèlo , o

Gouyon qrhsimopoie ma pleiˆda qarakthristik¸n perigraf  rujmoÔ [36℄. Qrhsimopoi¸nta

qarakthristikˆ perigraf  tempo , periodikìthta , isqÔo periodikìthta , kèntrou bˆrou i-

stogrˆmmato periodikìthta , qarakthristikˆ perigraf  kroustìthta ( per ussiveness ), kai

qarakthristikˆ perigraf  tou istogrˆmmato paÔsewn ( interval histogram ).

3.3.3 Armonikˆ kai melwdikˆ qarakthristikˆ

H armona mpore na oriste w h qr sh tou mousikoÔ tìnou maz me ti armonikè (sugqo-

de sthn perptwsh th dutik  mousik  ). Antijètw , melwda orzoume w ma akolouja

jemeliwd¸n suqnot twn. En¸ h armona anafèretai sun jw w to kˆjeto stoiqeo th mou-

sik  , h melwda anafèretai w to orizìntio stoiqeo th . H melwdik  kai armonik  anˆlush

qrhsimopoietai ai¸ne apì mousikolìgou gia th melèth mousik¸n dom¸n kai h qr sh th sto

prìblhma th anagn¸rish mousikoÔ edou mpore na odhg sei se jetikˆ apotelèsmata. O

Gomez kˆnei ma anaskìphsh th perigraf  kai exagwg  melwda apì hqhtikì s ma sto

[34℄.

O Tzanetˆkh to 2002 apopeirˆjhke pr¸to na qrhsimopoi sei qarakthristikˆ basismè-

na ston mousikì tìno gia anagn¸rish mousikoÔ edou [84℄. To diˆnusma qarakthristik¸n

39
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

mousikoÔ tìnou ( pit h ontent feature set ) pou protenei baszetai se pollaplè teqnikè

anqneush mousikoÔ tìnou. Sthn teqnik  pou protenetai, to s ma aposuntjetai se dÔo

suqnotikè mpˆnte , pˆnw kai kˆtw twn 1000 Hz . Upologzontai sthn sunèqeia oi peribˆl-

louse gia kˆje mpˆnta. Ajrozontai oi peribˆllouse kai upologzetai h autosusqètish

tou apotelèsmato . Oi korufè th autosusqètish antistoiqoÔn stou mousikoÔ tìnou

tou s mato . Sth sunèqeia dhmiourgetai èna istìgramma mousikoÔ tìnou ( pit h histogram )

me bˆsh ti korufè . Oi suqnìthte pou antistoiqoÔn se kˆje koruf  tou istogrˆmmato

metatrèpontai se mousikè nìte kai onomˆzontai sÔmfwna me to prwtìkollo MIDI . Dhmiour-

goÔntai dÔo istogrˆmmata, to diplwmèno ( folded ), pou èqei uposte thn metatrop  se MIDI ,

kai to aplwmèno ( unfolded ), pou den èqei uposte thn metatrop . Ta telikˆ qarakthristikˆ

pou prokÔptoun apì to istìgramma mousikoÔ tìnou enai:

 FA0 : plˆto th mègisth koruf  tou folded istogrˆmmato , antistoiq¸nta sthn

tonik  klmaka tou degmato .

 UP0 : Perodo th mègisth koruf  sto unfolded istìgramma, pou antistoiqe sto

eÔro se oktˆbe tou èrgou.

 FP0 : Perodo th mègisth koruf  tou folded istogrˆmmato .

 IPO1 : Diˆsthma mousik¸n tìnwn anˆmesa sti dÔo uyhlìtere korufè tou folded
istogrˆmmato , deqonta to kÔrio tonikì diˆsthma pou qrhsimopoietai sto kommˆti

(sta aplˆ èrga isqÔei to diˆsthma tonik  -despìzousa ).

 SUM : to ˆjroisma tou istogrˆmmato .

O Cataltepe epsh qrhsimopoie 5 qarakthristikˆ basismèna sto istìgramma mousikoÔ

tìnou [20℄.

3.4 Algìrijmoi Anagn¸rish

Oi teqnikè anagn¸rish mousikoÔ edou katatˆssontai se dÔo kathgore sth bibliografa:

sti mejìdou qwr epbleyh (ìpou den dnetai ierarqa, allˆ h ierarqa prokÔptei omadopoi¸-

nta ta mousikˆ kommˆtia) kai sti mejìdou me epbleyh (ìpou dnetai h ierarqa mousik¸n

eid¸n, to sÔsthma ekpaideÔetai kai sth sunèqeia elègqetai h apìdosh tou sust mato me ta

dedomèna elègqou).

3.4.1 Mèjodoi qwr epbleyh

Sti mejìdou qwr epbleyh, ta dedomèna omadopoioÔntai me bˆsh ta qarakthristikˆ tou

kai ta mètra omoiìthta pou qrhsimopoioÔntai, kai prokÔptei h taxinìmhsh. To pleonèkthma

40
3.4. ALGŸ
ORIJMOI ANAGNŸ
WRISHS

aut¸n twn mejìdwn enai ìti den upˆrqei o periorismì ma dedomènh ierarqa , pou mpore na

perièqei asˆfeie kai epikalÔyei . Epsh , kˆpoia kommˆtia polÔ aplˆ mpore na mhn an koun

se kanèna apì ta dosmèna mousikˆ edh.

Kˆje mousikì kommˆti anaparistˆtai apì èna sÔnolo qarakthristik¸n, ìpw parousiˆ-

sthke sthn enìthta 3.3. Epilègetai katìpin èna mètro omoiìthta gia na sugkrnei ta kom-

mˆtia metaxÔ tou . Oi algìrijmoi omadopohsh qrhsimopoioÔn ta mètra omoiìthta gia thn

orgˆnwsh twn mousik¸n kommati¸n se omˆde pou èqoun koinˆ qarakthristikˆ.

To pio aplì mètro omoiìthta pou qrhsimopoietai gia na metr sei thn apìstash anˆmesa

se dÔo dianÔsmata qarakthristik¸n enai ete h Eukledia apìstash   to mètro omoiìthta

sunhmitìnou ( osine similarity measure ). 'Omw , autˆ ta mètra omoiìthta prèpei na efar-

mìzontai mìno ìtan ta dianÔsmata qarakthristik¸n enai qronoametˆblhta. Alli¸ , dÔo

parìmoia mousikˆ kommˆtia mpore na jewrhjoÔn anìmoia apì to mètro an ta qarakthristikˆ

tou metablhjoÔn me ton qrìno. Gia na dhmiourghje ma qronoametˆblhth anaparˆstash

mia qronoseirˆ dianusmˆtwn qarakthristik¸n, sun jw dhmiourgoÔntai statistikˆ montèla

th katanom  twn qarakthristik¸n kai èpeita qrhsimopoietai kˆpoio mètro omoiìthta gia

thn sÔgkrish twn katanom¸n.

Tupikˆ montèla enai ta montèla megmato Gkaousian¸n ( Gaussian mixture models -


GMMs GMMs ). qrhsimopoi jhkan gia thn dhmiourga montèlwn qroiˆ sto [60℄. Epsh ,

Kullba k-Leibler
h apìklish qrhsimopoietai gia na upologsei thn apìstash anˆmesa se

katanomè allˆ den sust netai gia GMMs . 'Allo mètro pou qrhsimopoietai gia sÔgkrish

katanom¸n enai h apìstash Earth Movers , pou qrhsimopoi jhke apì ton Pampalk [67℄.

Epsh , o Shao qrhsimopohse krummèna montèla Markov hidden Markov models - HMMs
( )

gia na montelopoihje h sqèsh twn qarakthristik¸n ston qrìno [81℄.

'Oson aforˆ tou algorjmou omadopohsh pou qrhsimopoioÔntai, o Shao qrhsimopo-

hse ierarqikoÔ susswreutikoÔ algorjmou omadopohsh ( agglomerative hierar hi al lu-


stering) sto [81℄. Oi ierarqiko susswreutiko algìrijmoi arqzoun me N omˆde (ìpou N
enai o arijmì twn kommati¸n) kai en¸noun stadiakˆ ti omˆde qrhsimopoi¸nta èna mètro

omoiìthta . O auto-organwnìmeno qˆrth ( self-organizing map - SOM ) kai o auxanìmeno

ierarqikì auto-organwnìmeno qˆrth ( growing hierar hi al self-organizing map - GHSOM )

qrhsimopoioÔntai gia omadopohsh dedomènwn. Ta dedomèna anaparist¸ntai se ènan disdiˆ-

stato q¸ro, ètsi ¸ste ìmoia dianÔsmata qarakthristik¸n apeikonzontai kontˆ. Ta SOMs
enai teqnhtˆ neurwnikˆ dktua qwr epbleyh pou dhmiourgoÔn antistoiqe metaxÔ dedomè-

nwn poll¸n diastˆsewn kai se q¸rou qamhl¸n diastˆsewn. Ta GHSOM apoteloÔn eidik 

perptwsh twn SOM , ta opoa qrhsimopoioÔn ma ierrqik  dom  poll¸n epipèdwn, ìpou se

kˆje eppedo antistoiqe èna SOM . O Rauber qrhsimopohse GHSOMs gia na anaparast sei

optikˆ mousikè sullogè sto [74℄.

41
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

Ma efarmog  omadopohsh mousik¸n kommati¸n me endiafèron enai o Qˆrth tou Mo-
zart 7
 , pou dhmiourg jhke apì to Teqnikì Panepist mio Biènnh . Ston qˆrth omadopoioÔntai

ìla ta èrga tou W. A. Mozart se nhsiˆ mousik¸n eid¸n, ìpw sumfwne , sonˆte , kon-

sèrta, tragoÔdia, k.a. Sto Sq ma 3.3 apeikonzetai o Qˆrth , me diakritˆ ta nhsiˆ, me ti

omˆde mousik¸n eid¸n.

Sq ma 3.3: O Qˆrth tou Mozart .

To kÔrio meionèkthma twn mejìdwn qwr epbleyh enai ìti oi prokÔptouse omˆde den

onomatzontai. Genikˆ ìmw , oi omˆde den apeikonzoun aparathta kˆpoio mousikì edo ,

allˆ mousikˆ kommˆtia me kˆpoia ìmoia qarakthristikˆ. Uposthrzetai ìti h ènnoia tou

mousikoÔ edou mpore na qaje se bˆjo qrìnou kai na prokÔyei ma orgˆnwsh twn mousik¸n

dedomènwn me bˆsh thn omoiìthta sto [76℄.

3.4.2 Mèjodoi me epbleyh

Oi mèjodoi anagn¸rish mousikoÔ edou me epbleyh èqoun melethje perissìtero sthn bi-

bliografa. Se autè ti teqnikè upˆrqei ma bˆsh mousik¸n kommati¸n, h opoa prèpei

na antistoiqhje se ma dosmènh ierarqa, qrhsimopoi¸nta algorjmou mhqanik  mˆjhsh .

Sto pr¸to b ma, to sÔsthma ekpaideÔetai me dedomèna gia ta opoa enai gnwstì to mousikì

edo . Sto deÔtero b ma, taxinomoÔntai sto sÔsthma arqea gia ta opoa den enai gnwstì to

edo (arqea elègqou). Parakˆtw perigrˆfontai oi sun jei taxinomhtè pou efarmìzontai

sto prìblhma anagn¸rish mousikoÔ edou :

7 http://www.ifs.tuwien.a .at/mir/mozart/

42
3.4. ALGŸ
ORIJMOI ANAGNŸ
WRISHS

 Taxinomht  K -plhsièsterwn geitìnwn ( K -nearest neighbor lassi er - KNN): mh


parametrikì taxinomht  pou baszetai sthn idèa ìti èna mikrì arijmì geitìnwn

kajorzei thn apìfash gia to dedomèno elègqou. Dosmènou enì dianÔsmato qarakth-

ristik¸n, epilègontai ta K plhsièstera dianÔsmata se autì. To diˆnusma an kei sthn

klˆsh pou sunantˆtai perissìtero anˆmesa sta K plhsièstera dianÔsmata. O KNN


algìrijmo qrhsimopoi jhke apì ton Tzanetˆkh [84℄ kai ton Pampalk [67℄.

 Montèla megmato Gkaousian¸n (Gaussian mixture models - GMMs ): montelo-

poioÔn thn katanom  twn dianusmˆtwn qarakthristik¸n. Sun jw qrhsimopoietai o

algìrijmo megistopohsh kai anamenìmenh tim  ( expe tation maximization - EM )

gia thn ektmhsh twn paramètrwn twn Gkaousian¸n. Ta GMMs sto pedo th anagn¸-

rish mousikoÔ edou qrhsimopoioÔntai sun jw gia thn dhmiourga montèlwn qroiˆ .

San taxinomhtè , mporoÔn na qrhsimopoihjoÔn se sunduasmì me èna krit rio mègisth

pijanofˆneia , gia thn eÔresh tou montèlou pou perissìtero tairiˆzei sto dosmèno kom-

mˆti elègqou. O Tzanetˆkh qrhsimopohse GMMs gia thn montelopohsh mousik¸n

eid¸n [84℄. O Burred qrhsimopohse ma dendroeid  dom  apì GMMs gia th monte-

lopohsh th mousik  ierarqa [19℄. Tèlo , o West qrhsimopoie ènan Gkaousianì

taxinomht  se dendroeid  dom  pou qrhsimopoie thn apìstash Mahalanobis .

 Krummèna montèla Markov (hidden Markov models - HMMs 8 ) : qrhsimopoioÔntai

kurw sthn anagn¸rish omila , lìgw th ikanìthtˆ tou sto qeirismì qronoseir¸n.

Qrhsimopoi jhkan apì ton S aringella [77℄ gia anagn¸rish se 7 mousikˆ edh kai apì

ton Soltau [82℄ gia anagn¸rish se 4 mousikˆ edh.

 Anˆlush GrammikoÔ DiaqwrismoÔ ( linear dis riminant analysis - LDA ): h basik 

idèa enai h eÔresh enì grammikoÔ metasqhmatismoÔ pou diaqwrzei ìso to dunatìn

kalÔtera ti klˆsei . O West qrhsimopohse LDA gia na mei¸sei ti diastˆsei sto

prìblhma th taxinìmhsh prin na montelopoi sei ta dedomèna tou me Gkaousian  ka-

tanom  [88℄.

 Mhqanè Edrawn Dianusmˆtwn (support ve tor ma hines - SVMs ): majanoun

to upereppedo mègistou perijwrou ( maximum margin hyperplane ), kajist¸nta tou

anjektikoÔ sthn uperprosarmog  ( over- tting ). Qrhsimopoi jhkan apì ton S arin-


gella [77℄ kai ton Lidy [58℄ gia peirˆmata anagn¸rish mousikoÔ edou . O Mandel
qrhsimopohse SVMs se sunduasmì me thn apìstash Kullba k-Leibler [60℄.

8 'Ena eisagwgikì odhgì gia HMMs upˆrqei sto:


http://statwww.epfl. h/tea hing/e oleDo torale/Katz/hmmtut.pdf

43
KEFŸ
ALAIO 3. ANAGNŸ
WRISH MOUSIKOŸ
U EŸ
IDOUS

 Teqnhtˆ neurwnikˆ dktua (arti ial neural networks - ANN ): qrhsimopoioÔn san

domikˆ stoiqea teqnhtoÔ neur¸ne gia thn eplush problhmˆtwn. Ta pio diadedomèna

dktua enai ta polustrwmatikˆ per eptrons multilayer per eptrons - MLPs


( ), ta opoa

mporoÔn na proseggsoun opoiad pote mh grammik  sunˆrthsh. O Soltau qrhsimopohse

ma parallag  twn MLPs gia anagn¸rish mousikoÔ edou , ta opoa periègraye w

rht  qronik  montelopohsh me neurwnikˆ dktua ( expli it time modeling with neural
networks - ETM-NN ) [82℄. Sto ETM-NN , kˆje krummèno neur¸na mpore na jewrhje

w èna mousikì gegonì (mousik  frˆsh, rujmikì   melwdikì motbo).

3.5 Apotelèsmata - Telikˆ Sumperˆsmata

AkoloujoÔn sugkentrwtikˆ apotelèsmata gia ti suqnˆ qrhsimopoioÔmene bˆsei mousik¸n

eid¸n pou qrhsimopoioÔntai sth bibliografa. Gia to sÔnolo dedomènwn GTZAN , 61.0% enai

h mègisth akrbeia tou Tzanetˆkh [84℄. O Lidy me SVMs anèfere 74.9% akrbeia [58℄, en¸

o Li 78.5%. Gia to sÔnolo ISMIRrhythm , h akrbeia sto [58℄  tan 84.2%, sto [36℄ 90.1%,

kai sto [25℄ 96.0%. Gia to sÔnolo dedomènwn ISMIRgenre , h akrbeia tou Lidy enai 70.4%,

tou Tzanetˆkh 71.3%, tou West 78.3%, kai tou Pampalk 82.3% [58℄. Gia to sÔnolo ISMIR
2005 , thn kalÔterh akrbeia eqan oi Bergstra, Casagrande, kai E k, me 82.34% kai gia ta 2

sÔnola dedomènwn.

Genikˆ, to pedo th anagn¸rish mousikoÔ edou enai sqetikˆ nèo, kaj¸ oi pr¸te a-

xiìloge ergase pragmatopoi jhkan apì to 1998 kai èpeita. Parìla autˆ, gnwrzei megˆlh

ˆnjhsh kai mpore na jewrhje ìti enai to pio energì pedo ston q¸ro th epexergasa  qou

( audio pro essing ), qwr na sumperilhfje h epexergasa omila . Ta sÔnola dedomènwn pou

upˆrqoun sthn bibliografa apoteloÔn sthn pleioyhfa tou arketˆ realistikè peript¸sei ,

pou kalÔptoun toulˆqiston ti anˆgke taxinìmhsh pou antimetwpzei o mèso qr sth . To

prìblhma ìmw th eÔresh ma kajolik  taxinìmhsh , h opoa na mpore na qrhsimopoihje

apì etaire orgˆnwsh /p¸lhsh mousikoÔ ulikoÔ, enai akìma anoiktì. Ta qarakthristikˆ

pou èqoun protaje plèon kalÔptoun ti anˆgke gia autìmata sust mata anagn¸rish , ka-

j¸ plèon h akrbeia twn susthmˆtwn enai megalÔterh se sqèsh me thn akrbeia anagn¸rish

mousikoÔ edou apì mh eidikeumènou anjr¸pou .

44
Kefˆlaio 4

Paragontopohsh mh arnhtik¸n
tanust¸n

4.1 Eisagwg 

Kentrikì stìqo twn pedwn th mhqanik  mˆjhsh kai anagn¸rish protÔpwn enai h eÔre-

sh katˆllhlh anaparˆstash twn dedomènwn. Shmantikì rìlo èqoun epsh kai oi teqnikè

anˆlush upoq¸rwn ( subspa e analysis te hniques ), oi opoe apokalÔptoun domè qamhl¸n

diastˆsewn se q¸rou poll¸n diastˆsewn. Ma katˆllhlh anaparˆstash twn dedomènwn

se q¸ro qamhl¸n diastˆsewn mpore na odhg sei sthn eÔresh qarakthristik¸n gnwrismˆ-

twn twn dedomènwn, na kˆnei pio eÔkolh thn ermhnea tou , kai na gnei kalÔtera anˆlush

twn paragìntwn pou sunjètoun ta dedomèna kajeautˆ. Upˆrqoun dÔo stìqoi se ma tètoia

anaparˆstash: pr¸ton, na upˆrqei ermhnea twn apotelesmˆtwn (sthn perptwsh teqnik¸n

anˆlush upoq¸rwn na ermhneÔetai h anaparˆstash qamhloÔ arijmoÔ diastˆsewn) kai deÔ-

teron, na enai upologistik¸ efikt  h mèjodo anˆlush twn dedomènwn.

Sto parìn kefˆlaio, protenetai ma nèa polugrammik  mèjodo anˆlush upoq¸rwn, h

paragontopohsh mh arnhtik¸n tanust¸n ( non-negative tensor fa torization - NTF ). Sthn

enìthta 4.2 parousiˆzontai sunoptikˆ oi proteinìmene teqnikè anˆlush upoq¸rwn sthn

bibliografa, qwrismène se grammikè kai polugrammikè . Sthn enìthta 4.3 parousiˆzetai

analutikˆ h teqnik  th paragontopohsh mh arnhtik¸n pinˆkwn ( non-negative matrix fa to-


rization - NMF ), h opoa apotele ma exeidikeumènh ekdoq  th NTF gia tanustè 2h tˆxh ,

koin¸ pnake . Sthn enìthta 4.4 anafèrontai oi teqnikè paragontopohsh mh arnhtik¸n ta-

nust¸n pou protenontai sthn bibliografa. O proteinìmeno algìrijmo paragontopohsh

mh arnhtik¸n tanust¸n qrhsimopoi¸nta apoklsei Bregman parousiˆzetai sthn enìthta

4.5. Sthn enìthta 4.6 parousiˆzetai o algìrijmo gia sugkekrimène apoklsei Bregman .

Sthn enìthta 4.7 parousiˆzontai algìrijmoi NMF gia tanustè 3 diastˆsewn. O algìrijmo

45
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

taxinìmhsh NTF me epbleyh dnetai sthn enìthta 4.8. Tèlo , kˆpoia genikˆ sumperˆsmata

gia ton NTF algìrijmo dnontai sthn enìthta 4.9.

4.2 Teqnikè Anˆlush Upoq¸rwn

4.2.1 Grammikè Teqnikè Anˆlush Upoq¸rwn

Sth bibliografa èqoun protaje pollè teqnikè anˆlush upoq¸rwn. H pio diadedomènh enai

h anˆlush prwteuous¸n sunistws¸n ( prin ipal omponent analysis - PCA ), h opoa meta-

sqhmatzei ta dedomèna se èna nèo sÔsthma suntetagmènwn, ètsi ¸ste na megistopoietai h me-

tablhtìthta ston kÔrio ˆxona (prwteÔousa sunist¸sa). Skopì th anˆlush anexˆrthtwn

sunistws¸n ( independent omponent analysis - ICA ) enai h anˆlush enì polumetablhtoÔ

s mato se prosjetikè sunist¸se , jewr¸nta san periorismì thn statistik  anexarthsa

twn sunistws¸n [46℄. H anˆlush paragìntwn ( fa tor analysis


) jewre èna grammikì montè-

lo mexh parìmoio me thn ICA , ètsi ¸ste oi parathroÔmene metablhtè twn dedomènwn na

dhmiourgoÔntai w grammikì sunduasmì mh parathroÔmenwn metablht¸n, pou onomˆzontai

parˆgonte . H lanjˆnousa shmasiologik  anˆlush ( latent semanti analysis - LSA ) [24℄ kai

h pijanotik  lanjˆnousa shmasiologik  anˆlush ( probabilisti latent semanti analysis -


PLSA ) [43℄ enai statistikè teqnikè pou prospajoÔn na antistoiqsoun dianÔsmata poll¸n

diastˆsewn se q¸ro qamhl¸n diastˆsewn kai qrhsimopoioÔntai kurw sta peda anˆkthsh

plhrofora kai epexergasa fusik  gl¸ssa . Tèlo , h paragontopohsh mh arnhtik¸n pi-

nˆkwn ( non-negative matrix fa torization - NMF ), h opoa ja analuje sto parìn kefˆlaio,

h opoa brskei thn paragontopohsh enì pnaka, me ton periorismì ìti perièqei mh arnhtikè

timè [55, 56, 45℄. Na shmeiwje ìti h PLSA lÔnei to prìblhma th NMF qrhsimopoi¸nta

apìklish Kullba k-Leibler [32℄.

4.2.2 Polugrammikè Teqnikè Anˆlush Upoq¸rwn

Antstoiqa me ti teqnikè anˆlush upoq¸rwn gia polumetablhtˆ dedomèna, prìsfata èqoun

protaje kai teqnikè anˆlush poludiˆstatwn polumetablht¸n dedomènwn ( multidimensio-


nal multivariate data analysis te hniques ). Koin¸ , en¸ oi teqnikè anˆlush upoq¸rwn gia

polumetablhtˆ dedomèna qrhsimopoioÔsan pnake kai teqnikè grammik  ˆlgebra gia thn

pragmatopohsh twn upologism¸n, oi teqnikè anˆlush polumetablht¸n dedomènwn qrhsi-

mopoioÔn tanustè kai teqnikè polugrammik  ˆlgebra gia thn pragmatopohsh twn upolo-

gism¸n. H anˆgkh gia thn dhmiourga twn poludiˆstatwn teqnik¸n pro lje apì ta peda th

yuqologa kai qhmea , ìpou èginan oi pr¸te apìpeire anˆlush kai ermhnea trisdiˆsta-

twn dedomènwn. To 1964 o Tu ker prìteine ma parallag  th PCA gia anˆlush pragmatik¸n

46
4.2. TEQNIKŸ
ES ANŸ
ALUSHS UPOQŸ
WRWN

trisdiˆstatwn pinˆkwn sto pedo th yuqometra . To montèlo pou anaptÔqjhke onomˆsthke

Tu ker3 model [85℄.

4.2.3 PARAFAC

H pio diadedomènh teqnik  anˆlush trisdiˆstatwn dedomènwn enai h PARAFAC (apì to

parallel fa tor analysis ), h opoa onomˆzetai epsh kai CANDECOMP anoni al


(apì to

de omposition PARAFAC
) [17℄. H protˆjhke gia ta peda th yuqometra ( psy hometri-
s hemometri s
) kai th qhmeiometra ( ). Basikì skopì th mejìdou PARAFAC enai h

aposÔnjesh enì trisdiˆstatou tanust  X 2 R I I I 1 2 3


se ˆjroisma exwterik¸n ginomènwn

dianusmˆtwn:
k
X
X= aj
bj
j (4.1)

j =1
H aposÔnjesh pragmatopoietai me algìrijmo enalassìmenwn elaqstwn tetrag¸nwn ( alter-
nating least squares - ALS ). Sthn PARAFAC lÔnetai to akìloujo prìblhma beltistopohsh :


X k
X
2

min
a;b;


Xi i i
1 2 3
aj i
( 1)

bj i

( 2) j (i3 ) (4.2)

i1 i2 i3 j =1
Basikì pleonèkthma th PARAFAC enai ìti brskei monadik  lÔsh me katˆllhlh epilog 

tou k. Sto Sq ma 4.1 apeikonzetai h leitourga th mejìdou PARAFAC (sto sugkekrimèno

montèlo eisˆgetai kai jìrubo , o opoo upologzetai w to lˆjo upologismoÔ).

Sq ma 4.1: Anaparˆstash th PARAFAC gia k = 2 (apì to [17℄).

4.2.4 Polugrammik  anˆlush sunistws¸n

O Kroonenberg èjese pr¸to ta jemèlia gia ènan algìrijmo pou na apotele polugrammik 

epèktash th anˆlush prwteuous¸n sunistws¸n [50℄. To montèlo tri¸n diastˆsewn pou

protenei enai basismèno sto montèlo tou Tu ker kai enai to akìloujo:

K X
X M
L X
Gi i i =
1 2 3
gi k hi l ei m klm
1 2 3 (4.3)

k=1 l=1 m=1

47
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

ìpou G 2 R I I I
1 2 3
kai C 2 R K LM . To montèlo mpore enallaktikˆ na grafe w :

G =C G H E 1 2 3 (4.4)

ìpou G2 R I1 K , H2 R I 2 L , kai E2 R I3 M . Gia thn eplush tou probl mato qrh-

simopoietai h mèjodo enallassìmenwn elaqstwn tetrag¸nwn ( alternating least squares -


ALS ).

To 1986, o Kapteyn kˆnei ma epèktash tou montèlou 3-mode omponents analysis pou

pr¸to eqe protenei o Tu ker , se N diastˆsei [47℄. Gia thn eplush tou probl mato

qrhsimopoietai h mejodo ALS .

O Lathauwer epsh protenei ma mèjodo polugrammik  aposÔnjesh idiazous¸n tim¸n

( singular value de omposition - SVD ) [51℄ [53℄. Onomˆzei to montèlo w SVD uyhl  tˆxh

( higher-order SVD - HOSVD ) kai enai to akìloujo:

G = S  U  U    N U N
1
(1)
2
(2) ( )
(4.5)

ìpou oi tanustè G; S 2 R I1 I2 IN kai oi pnake Ui 2( )


R Ii Ii . Oi pnake Ui( )
enai

orjog¸nioi. O kentrikì tanust  S enai ìlo-orjog¸nio (bl. Enìthta 2.3.7) kai upˆrqei

diˆtaxh twn upo-tanust¸n tou S , se analoga me thn anˆlush idiazous¸n tim¸n:

jjSin jj > jjSin jj >    > jjSin In jj > 0


=1 =2 = (4.6)

O Lathauwer apodeiknÔei ìti gia thn eÔresh tou HOSVD apaitetai upologismì SVD sta

anaptÔgmata tou G , en¸ o kentrikì tanust  S upologzetai sthn sunèqeia:

S=G U 1
(1)
T
 U 2
(2)
T
   N U N T :( )
(4.7)

ApodeiknÔetai ìti h lÔsh pou dnei h HOSVD enai monadik . Epsh , gia N = 2 o algìrijmo
enai dio me ton SVD . H HOSVD mpore na qrhsimopoihje akìma gia ton upologismì tou

n-bajmoÔ tou G. Tèlo , h efarmog  th HOSVD se ènan summetrikì anˆ dÔo tanust  G2
R I I I isoduname me thn anˆlush idiotim¸n ston G . Koin¸ , protenetai ma genikeumènh

ekdoq  th PCA se pollè diastˆsei .

4.2.5 Polugrammik  ICA

O Be kmann prìteine ma epèktash th pijanotik  anˆlush anexˆrthtwn sunistws¸n ( pro-


babilisti independent omponent analysis - PICA ) gia trei diastˆsei [6℄. H teqnik  pro l-

je apì thn PARAFAC kai qrhsimopoi jhke sthn anˆlush FMRI . Jewre èna montèlo ìmoio

me th PARAFAC sthn (4.1), sun ènan ìro GkaousianoÔ jorÔbou. Sto montèlo PICA , o ìro

48
4.2. TEQNIKŸ
ES ANŸ
ALUSHS UPOQŸ
WRWN

C A jewretai w o pnaka mexh (ìpw upˆrqei sthn ICA). To sÔmbolo sumbolzei to


ginìmeno Khatri-Rao , to opoo enai ginìmeno Krone ker anˆ st le enì pnaka:

A B = [a
b j   jaN
bN ℄
1 1 (4.8)

Praktikˆ ìmw to montèlo Tensor PICA den axiopoie thn plhrofora twn 3 diastˆsewn,

allˆ pragmatopoie ma parallag  th ICA sto anˆptugma tou tanust .

O Terzìpoulo protenei ma ˆllh epèktash th ICA , me ttlo Multilinear ICA [86℄.

Sto montèlo th polugrammik  ICA pou protenoun, pragmatopoietai ma aposÔnjesh tou

tanust  dedomènwn:

G = S  U  U    N U N
1
(1)
2
(2) ( )
(4.9)

ìpou o S enai o kentrikì tanust  mexh , kai oi pnake Ui ( )


enai oi anexˆrthte suni-

st¸se gia thn i-ost  diˆstash. Sthn ousa ìmw , to montèlo pou protenetai den èkfrˆzei

polugrammik  mexh (pq. ta poludiˆstata dedomèna na ekfrˆzontai w mexh tanust¸n),

allˆ pollaplasiasmì pinˆkwn. Epiprìsjeta, o proteinìmeno algìrijmo qrhsimopoie to

anˆptugma tou tanust  G gia kˆje diˆstash, ìpou pragmatopoie ICA gia thn eÔresh twn

anexˆrthtwn sunistws¸n. 'Ara pˆli den axiopoietai h poludiˆstath dom  tou tanust , allˆ

to prìblhma anˆgetai sthn eplush N grammik¸n problhmˆtwn. H teqnik  Multilinear ICA


qrhsimopoi jhke se tanustè tri¸n diastˆsewn gia anagn¸rish pros¸pwn.

O Yoo protenei ma epèktash th ICA , me ttlo Multiway ICA [89℄. H teqnik  baszetai

pˆnw sthn Multiway PCA h opoa praktikˆ enai efarmog  th PCA pˆnw sta anaptÔgmata

tou tanust  dedomènwn, gia ìle ti diastˆsei . Sthn Multiway ICA , pragmatopoietai ICA
sta anaptÔgmata G (2) kai G(3) tou tanust  G (jewretai ìti h 1h diˆstash enai h diˆstash

twn dedomènwn). AkoloÔjw , parˆgontai oi dÔo pnake mexh gia ti ˆlle 2 diastˆsei

twn dedomènwn. H mèjodo protˆjhke gia anˆlush kai epexergasa lummˆtwn kai èqei ta dia

meionekt mata pou èqei kai h mèjodo tou Terzìpoulou [86℄. Sto Sq ma 4.2 fanontai ta

anaptÔgmata tou arqikoÔ tanust  pou qrhsimopoie o Yoo gia ton upologismì th Multiway
ICA .

Tèlo , o He protenei sunduasmì polugrammik  anˆlush prwteuous¸n sunistws¸n kai

anˆlush anexˆrthtwn sunistws¸n [40℄. Protenetai h efarmog  polugrammik  PCA kai

katìpin efarmog  ICA sti asusqètiste metablhtè pou parˆgei h polugrammik  PCA , ètsi

¸ste ta apotelèsmata na enai anexˆrthta. H mèjodo pou protenetai èqei ta meionekt mata

th polugrammik  PCA se sunduasmì me ta meionekt mata th efarmog  grammik  ICA se

anˆptugma tanust .

49
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

Sq ma 4.2: AnaptÔgmata arqikoÔ tanust  (apì to [89℄).

4.3 Paragontopohsh mh arnhtik¸n pinˆkwn

4.3.1 Montèlo

Oi algìrijmoi paragontopohsh mh arnhtik¸n pinˆkwn ( non-negative matrix fa torization -


NMF ) an koun sti teqnikè anˆlush upoq¸rwn. Brskoun th mh arnhtik  paragontopohsh

enì dosmènou mh arnhtikoÔ pnaka. V diastˆsewn n  m,


An doje èna pnaka oi NMF
algìrijmoi dnoun san èxodo tou mh arnhtikoÔ pnake W kai H ètsi ¸ste:

V  WH (4.10)

ìpou o pnaka W èqei diastˆsei n  r, kai o pnaka H èqei diastˆsei r  m, me ton


periorismì (n + m)r < nm, ètsi ¸ste oi pnake W kai H na enai mikrìterh diˆstash apì

ton arqikì pnaka V, pou èqei w apotèlesma mia sumpiesmènh ekdoq  tou arqikoÔ pnaka

dedomènwn.

O NMF baszetai sthn sumpesh topik¸n kai ìqi olistik¸n qarakthristik¸n enì s mato ,

se sÔgkrish me ˆlle mejìdou aposÔnjesh dedomènwn, ìpw h PCA kai h ICA . Epsh

diafèrei se sqèsh me ti ˆlle mejìdou ìson aforˆ ton periorismì th mh arnhtikìthta

twn pinˆkwn. Autì ermhneÔetai ìti h plhrofora apojhkeÔetai san grammikì sunduasmì

basik¸n stoiqewn plhrofora (sunart sei bˆsh ), kai ìti prosjetiko sunduasmo poll¸n

50
4.3. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN PINŸ
AKWN

sunart sewn bˆsh anaparistoÔn thn plhrofora. Koin¸ o NMF baszetai sth jewra ìti

h antlhyh tou ìlou baszetai sthn antlhyh twn mer¸n tou, kˆti pou èrqetai se antjesh

me thn olistik  je¸rhsh th PCA , ìpou h plhrofora anaparstatai san sunduasmì ìlwn

twn sunart sewn bˆsh (idiodianÔsmata).

H (4.10) mpore na xanagrafte anˆ st le , ètsi ¸ste vi  Whi , i = 1; 2:::m, ìpou vi
kai hi enai oi antstoiqe st le twn V kai H. Me ˆlla lìgia, kˆje diˆnusma plhrofora

vi anaparstatai san grammikì sunduasmì twn sthl¸n tou W, zugismèno me ta stoiqea


tou H. An jewr soume to r san arijmì klˆsewn, to stoiqeo wij anaparistˆ ton bajmì me

ton opoon to i-ostì qarakthristikì an kei sthn j -ost  klˆsh dedomènwn, j = 1; 2; : : : ; r ,

en¸ to stoiqeo hjk deqnei ton bajmì me ton opoo to diˆnusma plhrofora vi an kei sthn

klˆsh j .

4.3.2 Basikì NMF (Standard NMF)

Gia thn paragontopohsh NMF qrhsimopoioÔntai algìrijmoi pou baszontai se epanalhptikè

antikatastˆsei ( update rules ) stou pnake W kai H. Gia na breje mia proseggistik 

paragontopohsh tou pnaka V prèpei na oristoÔn sunart sei kìstou pou orzoun thn

poiìthta th prosèggish . Ma tètoia sunˆrthsh mpore na kataskeuaste qrhsimopoi¸nta

apostˆsei metaxÔ dÔo mh arnhtik¸n pinˆkwn A = [aij ℄


B = [bij ℄. kai 'Ena qr simo mètro

enai to tetrˆgwno th Eukledeia apìstash metaxÔ twn A kai B:

X
jjA Bjj = 2
(aij bij ) 2
(4.11)

ij

h opoa èqei san kˆtw ìrio to mhdèn kai emfan¸ mhdenzetai ìtan A = B. 'Allo èna qr simo

mètro enai h Kullba k-Leibler apìklish (KL divergen e ) twn pinˆkwn:

 
X a
DKL (AjjB) = aij log ij aij + bij (4.12)

ij
bij

'Opw kai h Eukledeia apìstash, èqei san kˆtw ìrio to mhdèn, kai mhdenzetai ìtan A = B.
Den enai jewretai apìstash, giat den enai summetrik , kai anafèretai w apìklish tou A

apì to B. 'Ara jewroÔme dÔo diaforetikè ulopoi sei th mejìdou NMF san probl mata

beltistopohsh : H pr¸th enai h elaqistopohsh tou jjV WHjj se sqèsh me ta W kai H,2

me ton periorismì W; H > 0. To deÔtero prìblhma enai h elaqistopohsh tou D (VjjWH)

se sqèsh me ta W kai H, me ton periorismì W; H > 0.

ApodeiknÔetai ìti oi parakˆtw epanalhptiko pollaplasiastiko kanìne sunduˆzoun eu-

kola ulopohsh kai taqÔthta gia thn lÔsh twn parapˆnw problhmˆtwn [55℄:

Prìblhma 1 : H Eukledeia apìstash jjV WHjj 2


enai mh auxanìmenh stou epanalh-

51
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

ptikoÔ kanìne :

h j h j
(WT V) j w w
(VHT )i
i i
(WT WH) j (WHHT )i
(4.13)

Prìblhma 2 : H apìklish Kullba k-Leibler DKL(VjjWH) enai mh auxanìmenh stou

epanalhptikoÔ kanìne :

P wi vij P h j vij
h j h j P
i WH ij
( )
wi wi P
WH ij
j ( )
(4.14)
k wk j h j

ìpou = 1; 2:::r; i = 1; 2:::n; j = 1; 2:::m. 'Oson aforˆ thn efarmog  tou basikoÔ NMF, oi
pnake W kai H mporoÔn na arqikopoihjoÔn me tuqae mh arnhtikè timè , kai apodeiknÔetai

[55℄ ìti o NMF ja sugklnei se topikì elˆqisto.

4.3.3 Topikì NMF (Lo alized NMF)

'Ena apì ta meionekt mata tou BasikoÔ NMF enai ìti oi bˆsei tou enai qwrikˆ olikè , en¸

ja  tan protimhtèo na  tan topikè . Gia autìn ton lìgo, o Li prìteine ma ˆllh mèjodo upo-

q¸rwn, thn onomazìmenh topik  paragontopohsh mh arnhtik¸n pinˆkwn ( lo al non-negative


matrix fa torization - LNMF ), gia ekmˆjhsh qwrikˆ topik¸n anaparastˆsewn protÔpwn [56℄.

Orzonta U = [uij ℄ = WT W; B = [bij ℄ = HHT , mia nèa sunˆrthsh apìstash prèpei na

elaqistopoihje kˆtw apì 3 sumplhrwmatikoÔ periorismoÔ :

1. Mia sunist¸sa bˆsh den ja prèpei na aposuntjetai peraitèrw se perissìtere suni-

st¸se , ètsi ¸ste na elaqistopoihje o arijmì twn sunart sewn bˆsh pou qreiˆzo-
P
ntai gia na anaparast soun ton pnaka V. Koin¸ prèpei
i uii = min, ìpou uij enai

to ij -ostì stoiqeo tou pnaka U.


2. Diaforetikè bˆsei ja prèpei na enai ìso to dunatìn pio orjog¸nie , kˆti pou epi-
P
bˆlletai me thn sunj kh:
i6=j uij = min .

P
3. Mìno ta stoiqea pou dnoun thn pio shmantik  plhrofora prèpei na krathjoÔn:
i bii =
max , ìpou bij enai to ij -ostì stoiqeo tou pnaka B.
O sunduasmì twn parapˆnw periorism¸n dnei thn parakˆtw apìklish pou qrhsimopoietai

san sunˆrthsh apìstash ston LNMF :

 
X v X X
DLNMF (VjjWH) = vij log ij vij + yij + uij bii ; yij = (WH)ij (4.15)

i;j
yij i;j i

ìpou ; stajerè . Ma paragontopohsh topikoÔ NMF orzetai w h elaqistopohsh th

(4.15). Ma topik  lÔsh sthn parapˆnw elaqistopohsh brsketai qrhsimopoi¸nta tou

52
4.3. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN PINŸ
AKWN

parakˆtw epanalhptikoÔ kanìne :


s
X wi
h j = h j vij P (4.16)

i wi h j

wi
P
j vij P hw ji h j w
wi = P ; wi = P i (4.17)
j h j i wi
O periorismì orjogwnikìthta moiˆzei me ton periorismì pou epibˆllei h mèjodo PCA sti

sunart sei bˆsh . Parìla ta topikˆ qarakthristikˆ tou ìmw , o LNMF den kwdikopoie

plhrofore diaqwrismoÔ gia probl mata taxinìmhsh .

4.3.4 Araiì NMF (Sparse NMF)

Oi Hu Zhang
kai parousasan mia ˆllh parallag  tou NMF me ìnoma Araiì NMF sparse
(

non-negative matrix fa torization - SNMF ) [45℄, h opoa qrhsimopoietai gia na susqetsei

kanìne analoga (ratio rules ). Basikì skeptikì psw apì ton Araiì NMF enai ìti en¸

h NMF mèjodo enai epituq  sthn paragontopohsh pinˆkwn, den jètei ìmw periorismoÔ

puknìthta twn dedomènwn stou pnake . San apotèlesma, den enai se jèsh na pragmatopoi-

 sei paragontopohsh se ènan pnaka V pou èqei topikˆ araiˆ qarakthristikˆ sta dedomèna

tou.

Me bˆsh thn arqik  NMF mèjodo, protˆjhke h akìloujh sunˆrthsh kìstou :


 
X v X
DSNMF (VjjY) = vij log ij vij + yij + khj kl (4.18)

i;j
yij j
ìpou hj = (h j ; h j ; :::hrj )T , j = 1; 2; :; m,
1 2 enai h antstoiqh st lh tou pnaka H, kai to 
enai jetik  stajerˆ apoktoÔmenh apì empeira. W khj kl orzetai h l-nìrma tou dianÔsmato
st lh tou pnaka H. To prìblhma th paragontopohsh me SNMF orzetai w :

min DSNMF (VjjWH) 8i; j : wij > 0; hij > 0; 8i : kwikl = 1


W;H
(4.19)

Mia lÔsh gia thn parapˆnw elaqistopohsh me periorismoÔ dnetai apì tou parakˆtw ka-

nìne epanˆlhyh :

P
P wi
i vij wi h j
P h j
j vij wi h j
P

h j = h j P (
wi = wi )
P
i wi + 
(4.20)
j h j
Gia na gnei h lÔsh monadik , apaitetai h l -nìrma tou dianÔsmato st lh ston pnaka W na

enai monadiaa. Epiplèon, o pnaka H prèpei na rujmiste katˆllhla:

w X
wi = P i h j = h j wi (4.21)
i wi i
ApodeiknÔetai ìti h (4.18) den auxˆnei kˆtw apì tou kanìne epanˆlhyh , kai h sÔgklish

th epanˆlhyh enai egguhmènh [45℄.

53
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

4.3.5 Taxinìmhsh NMF qwr epbleyh

H sun jh mèjodo gia taxinìmhsh dedomènwn stou upoq¸rou pou dhmiourgoÔn oi NMF
mèjodoi perigrˆfetai sto [7℄. V dhmiourgetai qrhsimopoi¸nta dedomèna apì to
O pnaka

sÔnolo ekpadeush (training set). Kˆje st lh tou V, vj perièqei èna diˆnusma qarakthri-

stik¸n. H diadikasa ekpadeush ekteletai efarmìzonta ènan algìrijmo NMF ston pnaka

dedomènwn, parˆgonta ètsi ton pnaka bˆsh W kai ton pnaka kwdikopohsh H.
Sthn fˆsh elègqou ( testing phase ), gia kˆje dedomèno elègqou to opoo anaparstatai me

èna diˆnusma qarakthristik¸n vtest , dhmiourgetai èna nèo diˆnusma elègqou-kwdikopohsh ,


me bˆsh thn parakˆtw sqèsh:

htest = Wy vtest (4.22)

ìpou Wy orzetai w o Moore-Penrose genikeumèno antstrofo (  yeudoantstrofo ) tou


W. Na shmeiwje ìti an o pnaka W èqei diastˆsei n  r, tìte o Wy ja èqei diastˆsei
r  n.
'Eqonta dhmiourg sei katˆ th diˆrkeia th ekpadeush N klˆsei apì dianÔsmata kw-

dikopohsh hl , l = 1; 2; : : : ; N , èna taxinomht  plhsièsterou getona efarmìzetai gia na

taxinom sei to degma elègqou vtest qrhsimopoi¸nta to mètro omoiìthta sunhmitìnou ( osine

similarity measure - CSM). To dedomèno elègqou katatˆssetai sthn klˆsh l0, ìpou:
 
hTtest hl
l0 = arg max
l=1;2;:::;N khtestkkhl k : (4.23)

Koin¸ , megistopoietai to sunhmtono th gwna metaxÔ twn htest kai hl . 'Ena enallaktikì

mètro mpore epsh na qrhsimopoihje, ìpou h klˆsh sthn opoa an kei to vtest kajorzetai
exetˆzonta kˆje stoiqeo tou htest :
l0 = arg max
i
hi;test (4.24)

ìpou hi;test enai to i-ostì stoiqeo tou htest .

4.3.6 Taxinìmhsh NMF me epbleyh

To basikì meionèkthma th taxinìmhsh me NMF algorjmou qwr epbleyh pou perigrˆfhke

sthn Enìthta 4.3.5 enai ìti sthn diadikasa ekpadeush den sumperilambˆnetai plhrofora

sqetikˆ me diaqwrismì klˆsewn sta dedomèna. Epiprìsjeta, oi arqikè tuqae timè twn

pinˆkwn W kai H mpore na ephreˆsoun thn sÔgklish tou algorjmou, kaj¸ h tim  th

antikeimenik  sunˆrthsh pou qrhsimopoietai apì ton ekˆstote NMF algìrijmo mpore na

egklwbiste se topikì elˆqisto, kai na odhg sei se esfalmènh paragontopohsh.

Sto [8℄ protenetai h dhmiourga enì taxinomht  me epbleyh, ìpou h diadikasa ekpa-

deush me NMF algorjmou ekteletai gia kˆje klˆsh dedomènwn xeqwristˆ. AkoloÔjw ,

54
4.3. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN PINŸ
AKWN

parˆgetai èna zèugo pinˆkwn W kai H gia kˆje klˆsh:


Vi = Wi Hi ; i = 1; 2;    ; N (4.25)

ìpou N enai o arijmì twn klˆsewn, Vi o pnaka dedomènwn gia thn klˆsh i, kai oi pnake
Wi ,Hi oi paragìmenoi mèsw NMF pnake gia thn klˆsh i. O arijmì twn sunistws¸n pou
qrhsimopoietai gia thn ekpadeush twn dedomènwn kˆje klˆsh dnetai apì:
 
ni mi
ri =
ni + mi
(4.26)

ìpou ni kai mi enai oi diastˆsei tou pnaka Vi . 'Omw , h bˆsh pou orzetai apì ti st le

tou pnaka Wi den enai orjokanonik . Gia na metatrape h bˆsh se orjokanonik , o protei-

nìmeno taxinomht  qrhsimopoie thn diadikasa orjogwniopohsh Gram-S hmidt ston Wi ,

efarmìzonta aposÔnjesh QR [9℄:

Wi = Qi Ri ; i = 1; 2;    ; N (4.27)

ìpou o pnaka Qi èqei diastˆsei n  r kai enai orjog¸nio . O pnaka Ri èqei diastˆsei
r  r kai enai ˆnw trigwnikì . Sunep¸ , o pnaka orjokanonik  bˆsh gia kˆje klˆsh enai
o Qi kai o nèo pnaka kwdikopohsh enai o:

H0i = Ri Hi ; i = 1; 2;    ; N: (4.28)

Katˆ th diˆrkeia th fˆsh elègqou, kˆje dedomèno elègqou anaparstatai apì to diˆnu-

sma qarakthristik¸n vtest . Sth sunèqeia, to vtest probˆlletai stou orjokanonikoÔ pnake
bˆsh Qi gia kˆje klˆsh, kai prokÔptei to diˆnusma kwdikopohsh -elègqou gia kˆje klˆsh:
0( )
i
htest = Qyi  vtest; (4.29)

ìpou Qyi enai o genikeumèno antstrofo tou Qi . Gia kˆje klˆsh, gnetai sÔgkrish tou
0 (i)
dianuÔsmato htest me ti st le tou pnaka H0i qrhsimopoi¸nta to mètro omoiìthta sunh-

mitìnou.

To diˆnusma pou megistopoie to CSM gia ton pnaka H0i qrhsimopoietai w mètro omoiì-
thta metaxÔ tou vtest kai th klˆsh i:
 0( )
iT i  0( )
htest hj
CSMi = j max
khtest kkh0j i k
0i (4.30)
; ;:::;r
=1 2 i ( ) ( )

0( )
ìpou to hj i j -ost  st lh tou pnaka H0i .
anaparistˆ thn Tèlo , to mègisto CSMi orzei se
poia klˆsh an kei to diˆnusma elègqou vtest :

l0 = arg i max
; ;:::;N
=1 2
fCSMig: (4.31)

'Ena diˆgramma th diadikasa elègqou qrhsimopoi¸nta taxinomht  NMF me epbleyh apei-

konzetai sto Sq ma 4.3.

55
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

0 (1)
Q 1
ht H0 1
CSM 1

0 (2)
vt Q 2
ht H0 2
CSM 2

arg
max
. . .

. . .

. . .

0(
QN ht N )
H0N CSMN

0( )
ht i = Qyi  vt

Sq ma 4.3: Diadikasa elègqou qrhsimopoi¸nta taxinomht  NMF me epbleyh (ta h0t kai vt
sumbolzoun ta h0test kai vtest , antstoiqa).

4.3.7 Epektˆsei mejìdwn NMF

O Donoho sto [26℄ jètei dÔo erwt mata pou sqetzontai me thn NMF mèjodo:

 Kˆtw apì poie sunj ke enai h paragontopohsh mh arnhtik¸n pinˆkwn kal¸ orismè-

nh, gia parˆdeigma pìte enai h paragontopohsh monadik ?

 Kˆtw apì poie sunj ke enai h paragontopohsh swst ?

ApodeiknÔei ìti h paragontopohsh enai monadik  gia dedomèna pou plhroÔn sugkekrimènou

periorismoÔ , anexart tw tou NMF algorjmou pou qrhsimopoietai. O periorismì enai

ìti prèpei na dhmiourgetai ma bˆsh dedomènwn ìpou ìla ta epijumhtˆ mèrh gia anqneush

ja prèpei na emfanzontai, me ìlou tou dunatoÔ sunduasmoÔ . Dhmiourg¸nta ma bˆsh

eikìnwn me basikˆ gewmetrikˆ sq mata pou na plhro ton periorismì (onomˆzetai Separable
Fa torial Arti ulation Families ), o Donoho deqnei ìti upˆrqei monadik  lÔsh. Se efarmogè

pragmatikoÔ kìsmou ìmw , den enai dunatìn na kataskeuaste ma tètoia bˆsh dedomènwn pou

na egguˆtai monadik  lÔsh, eidikˆ sthn perptwsh pou h NMF pragmatopoietai diereunhtikˆ,

mh gnwrzonta ta mèrh pou apoteloÔn ta dedomèna. To egqerhma kajstatai akìma pio

dÔskolo sthn efarmog  ˆllwn tÔpwn dedomènwn, ìpw hqhtik¸n eggraf¸n.

O Allbright protenei dÔo nèou algorjmou giaNMF qrhsimopoi¸nta thn teqnik  twn

enallassìmenwn elaqstwn tetrag¸nwn ( alternative least squares - ALS ) [1℄. O pr¸to algì-

rijmo onomˆzetai ACLS alternating onstrained least squares


( - enallassìmenwn elaqstwn

tetrag¸nwn me periorismoÔ ). LÔnei to akìloujo prìblhma:

h jjvj Whj jj + H jjhjj s:t: H > 0; hj > 0


min 2 2
2 2
(4.32)
j

56
4.3. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN PINŸ
AKWN

ìpou to H enai mètro arawsh . O ìro jjhjj2


2
qrhsimopoietai gia na proseggsei thn

araiìthta tou dianÔsmato h. O deÔtero algìrijmo antikajistˆ to jjhjj 2


2
me to mètro

spar(h), to opoo dnetai apì ton tÔpo:


pm jjhjj =jjhjj
spar(h) = pm 1 : 1 2
(4.33)

O deÔtero algìrijmo onomˆzetai AHCLS alternating Hoyer- onstrained least squares


( ),

kaj¸ o tÔpo gia thn araiìthta protˆjhke apì ton Hoyer [44℄. ApodeiknÔetai ìti oi algì-

rijmoi ACLS kaiAHCLS sugklnoun, allˆ endèqetai na pagideutoÔn se topikì elˆqisto   se

shmea sèlla ( saddle points ). 'Omw , stou dÔo parapˆnw algorjmou , h mh arnhtikìthta

epibˆlletai ( ad-ho non-negativity enfor ement ), jètonta ti arnhtikè timè twn pinˆkwn

W kai H se me mhdèn. Autì kajistˆ tou algorjmou mh elkustikoÔ apì jewrhtik 

skopiˆ.

O Allbright epsh jètei to er¸thma sqetikˆ me thn arqikopohsh tim¸n stou pnake W
kai H [1℄. Enai gnwstì ìti h sÔgklish twn algorjmwn NMF exartˆtai apì ti arqikè timè

twn pinˆkwn. Se kˆje perptwsh, ma orj  arqikopohsh mpore na aux sei thn taqÔthta

sÔgklish twn algorjmwn, kaj¸ kai thn akrbeiˆ tou . H tupik  arqikopohsh tim¸n gia

tou algorjmou NMF enai h arqikopohsh twn pinˆkwn W kai H me tuqae timè sto

diˆsthma [0; 1℄ . O Allbright uposthrzei ìti h tuqaa arqikopohsh pinˆkwn den odhge se

kal  poiìthta paragontopohsh (h poiìthta paragontopohsh orzetai me bˆsh kˆpoio

krit rio lˆjou anˆmesa ston pnaka V kai to ginìmeno WH), eidikˆ stou algorjmou

NMF pou qrhsimopoioÔn ALS teqnikè . Protenontai 6 diaforetikè arqikopoi sei , oi opoe

paratjentai ston Pnaka 4.1. H arqikopohsh Kèntrou SVD apaite efarmog  SVD ston

pnaka V. H Tuqaa A ol arqikopoie kˆje st lh tou W me bˆsh ton mèso ìro p tuqawn
sthl¸n tou V. H mèjodo Tuqaa C qrhsimopoie thn aposÔnjesh CUR [29℄. Tèlo , h

mèjodo Pnaka Sumpt¸sewn ( o-o uren e matrix ) qrhsimopoie ton pnaka C = VVT . Na

shmeiwje ìti oi algìrijmoi pou qrhsimopoioÔn ALS teqnikè apaitoÔn arqikopohsh mìno tou

pnaka W.

Pnaka 4.1: Mèjodoi arqikopohsh gia algorjmou NMF .

'Onoma Arqikopohsh Pleonekt mata Meionekt mata


Tuqaa Gr gorh Pukno pnake , sÔgklish se topikˆ elˆqista

Anamenìmenh Tim  Elatt¸nei # epanal yewn Megˆlo upologistikì kìsto

Kèntrou SVD Elatt¸nei # epanal yewn Prèpei na gnwrzoume parˆgonta SVD


Tuqaa A ol Gr gorh Mikr  mìno elˆttwsh epanal yewn

Tuqaa C Gr gorh Mh apotelesmatik 

Pnaka Sumpt¸sewn Elatt¸nei # epanal yewn Megˆlo upologistikì kìsto

57
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

Prosèggish Mh Arnhtik¸n Pinˆkwn


O Sra protenei ma enopohsh twn diafìrwn NMF algorjmwn pou èqoun protaje sthn bi-

bliografa [83℄. H enopohsh aut  gnetai upì to prsma ìti ìloi oi algìrijmoi prospajoÔn

na lÔsoun èna prìblhma beltistopohsh , sto opoo h sunˆrthsh kìstou enai ma apìklish

Bregman (bl. Enìthta 2.5). Prˆgmati, oi algìrijmoi pou prìteine o Lee [55℄ baszontai sthn

Frobenius nìrma kai sthn apìklish Kullba k-Leibler Bregman


, oi opoe enai apoklsei . H

nìrma Frobenius qrhsimopoietai epsh apì ton Hoyer [44℄, maz me ènan epiplèon ìro pou

ekfrˆzei thn araiìthta tou pnaka H Sra. O onomˆzei thn oikogèneia algorjmwn pou prote-

nei w Prosèggish Mh Arnhtik¸n Pinˆkwn ( non-negative matrix approximation - NNMA ,

jèlonta pio orjˆ na dhl¸sei ìti oi algìrijmoi brskoun prosèggish paragontopohsh , kai

ìqi thn paragontopohsh kajeaut ). Ta probl mata pou tjentai enai dÔo: (V2 Rm
+
n ,
W 2 R mr kai H 2 R rn ):

(1) Wmin
;H>
D (WH; V) + (W) + (H)
0
(4.34)

(2) Wmin
;H>
D (V; WH) + (W) + (H)
0
(4.35)

Oi sunart sei (W) (H), penalty fun tions


enai sunart sei poin  ( ), epibˆllonta e-

piprìsjetou periorismoÔ stou pnake W kai H. To pr¸to prìblhma pou jètei o Lee
[55℄ gia parˆdeigma, qrhsimopoie thn apìklish Bregman (x) = x me (W) = 0 1
2
2
, me kai

(H) = 0, kai antistoiqe sto prìblhma 2 tou Sra . Lee


To deÔtero prìblhma pou jètei o ,

qrhsimopoie thn (x) = x log(x), me (W) = 0 (H) = 0


kai , kai antistoiqe sto prìblhma

2 tou Sra . To prìblhma pou jètei o Hoyer (x) = x


[44℄, qrhsimopoie thn (W) = 0 1
2
2
me

kai (H) = 1T H1 .

Gia thn eplush tou probl mato 1 o Sra qrhsimopoie pollaplasiastikoÔ kanìne , qrh-

simopoi¸nta bohjhtikè sunart sei :

Orismì 4.3.1. Bohjhtik  Sunˆrthsh


( ) Ma sunˆrthsh G( ; 0 ) onomˆzetai bohjhtik 

sunˆrthsh gia thn F ( ) an:


(1) G( ; ) = F ( )

) > F ( ); 8 ~.
(2) G( ; ~

An h G( ; ~ ) enai bohjhtik  sunˆrthsh th F ( ), tìte h F den auxˆnei me ton epanalhptikì


kanìna:

t = arg min
+1

G( ; )
t (4.36)

Gia thn eÔresh epanalhptikoÔ kanìna gia thn eplush tou probl mato 1 w pro to h,
w F (h) jewretai h:
 
X X X
F (h) = ( wij hj ) (vi ) + (vi )((Wh)i vi ) ; (4.37)

i j i

58
4.4. PROTEINŸ
OMENES MŸ
EJODOI

ìpou () = 0() . Me bˆsh ton Orismì 4.3.1, dhmiourgetai h antstoiqh bohjhtik  sunˆrthsh:
   
w h
G(h; h~ ) =
X X
ij  ij j (vi ) + (vi )((Wh)i vi ) ; (4.38)

ij
ij i
ij = (wij h~ j )=( l wil h~ l ). Gia thn eÔresh epanalhptikoÔ kanìna gia to h, prèpei na
P
ìpou

elaqistopoihje h G(h; h ~ ) w pro to h. Autì pragmatopoietai lÔnonta thn exswsh


G
hp =0 . Gia thn eÔresh analutik  lÔsh , jewretai ìti h () enai diaqwrsimh, dhla-

d  (xy) = (x) (y) . Me bˆsh thn parapˆnw je¸rhsh, prokÔptei o pollaplasiastikì

kanìna enhmèrwsh gia to stoiqeo hp :


 

hp h~ p  1
[WT (v)℄p :
[WT (Wh~ )℄p
(4.39)

Antstoiqa prokÔptei kai o kanìna enhmèrwsh gia ta stoiqea tou pnaka W:


 

wp w~p  1
[ (vT )HT ℄p :
[ (w~ T H)HT ℄p
(4.40)

O Sra sth sunèqeia kalÔptei kai peript¸sei me mh mhdenikè sunart sei poin  , kaj¸

epsh peript¸sei me apoklsei ektì twn Bregman , ìpw oi apoklsei Csiszar kai Young .

4.4 Proteinìmene Mèjodoi

Oi pr¸toi algìrijmoi paragontopohsh jetik¸n pinˆkwn ( positive matrix fa torization ) pro-

tˆjhkan to 1994, kai oi pr¸toi algìrijmoi paragontopohsh mh arnhtik¸n pinˆkwn to 1999.

Apì to 2001 kai èpeita, èginan oi pr¸te apìpeire epèktash twn algorjmwn gia efarmo-

1
g  se tanustè tri¸n   parapˆnw diastˆsewn . Oi perissìtere protˆsei anafèrontai se

efarmogè epexergasa eikìnwn, ìpou qrhsimopoioÔntai tanustè 3 diastˆsewn.

O Welling to 2001 prìteine ma genkeush tou algorjmou jetik  paragontopohsh pinˆ-

kwn gia tanustè N -ost  tˆxh [87℄. O algìrijmo onomˆsthke paragontopohsh jetik¸n

tanust¸n ( positive tensor fa torization ). Sto montèlo pou protenetai, jewretai èna tanu-

I1 I2 IN
st  V2R +
. To montèlo ekfrˆzei ta stoiqea tou V w :

K
X
Vi i :::iN =
1 2
uj i uj i
1( 1) 2( 2)
   ujN iN ;
( )
(4.41)

j =1
ìpou uj i
1( 1)
enai to i 1 -ostì stoiqeo tou dianÔsmato uj .
1
O basikì skopì tou algorjmou

enai h elaqistopohsh tou lˆjou :

2 IN 
I1 IX K
X
2

RE = Vi i :::iN
1 2
uj
1(
j
i1 ) u2(i2 )  ujN (iN ) : (4.42)

i1 i2 :::iN j =1
1 Na shmeiwje ìti ta onìmata twn metablht¸n kai twn telest¸n pou qrhsimopoioÔntai sti proteinìmene

mejìdou den enai ta dia me autˆ pou qrhsimopoioÔntai sthn diatrib . Se ma apìpeira enopohsh th

bibliografa , paratjentai koinˆ sÔmbola gia metablhtè kai telestè .

59
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

Gia thn elaqistopohsh tou lˆjou , qrhsimopoietai o parakˆtw epanalhptikì pollaplasia-

stikì kanìna :

Ui ! Ui 
Vi i :::ii ii :::iN  M
1 2 1 +1

Ui  MT  M
(4.43)

PK j j j
ìpou M= j =1 u1(i1 ) u2(i2 ) : : : ui 1( ii 1)
uji +1(
j
ii+1 ) : : : uN (iN ) . Protenetai epsh kai ma sunˆr-

thsh kìstou pou apotele genkeush th sunˆrthsh pou protˆjhke apì ton Lee [55℄ (pa-

rallag  apìklish Kullba k-Leibler ). H ikanìthta tou algorjmou elègqetai se yhfiopoih-

mènou pnake tou zwgrˆfou Mondriaan , pou apoteloÔntai apì orjog¸nia se diaforetikè

diastˆsei kai qr¸mata (bl. Sq ma 4.4).

Sq ma 4.4: Pnake tou Mondariaan pou qrhsimopoi jhkan gia peirˆmata jetik  parago-

ntopohsh tanust¸n apì ton Welling [87℄.

To 2005 o Lim protenei ma polugrammik  parallag  tou NMF , h opoa ousiastikˆ

qrhsimopoie ton algìrijmo PARAFAC me epiplèon periorismì mh arnhtikìthta [59℄. Anafè-

retai sto prìblhma th eÔresh elˆqistou k gia to opoo ma tètoia aposÔnjesh enai efikt .
SÔmfwna me ton Lim , h idanik  paragontopohsh mh arnhtikoÔ tanust  dnetai apì to K

ìpou:
 K
X


arg min
K
jjV u
u


j
1
j
2
jj
ujN F ; uji 2 R Ii
+
: (4.44)

j =1
'Omw , den enai pˆnta efikt  h eÔresh tou idanikoÔ K, ˆra anazhtetai h mh idanik  lÔsh

sthn opoa, dojènto K, proseggzei ènan tanust  w ˆjroisma exwterik¸n ginomènwn dia-

nusmˆtwn. Tèlo , anafèretai ìti to prìblhma monadik  aposÔnjesh mh arnhtikoÔ tanust 

enai anoiktì, allˆ enai efiktì gia sugkekrimène morfè tanust¸n.

Oi Shashua kai Hazan prìteinan to 2005 ma genkeush tou algorjmou NMF gia tanu-

stè N -ost  tˆxh [80℄. H mèjodo onomˆsthke paragontopohsh mh arnhtik¸n tanust¸n

( non-negative tensor fa torization - NTF ). To proteinìmeno montèlo baszetai sthn idèa th

60
4.4. PROTEINŸ
OMENES MŸ
EJODOI

2
aposÔnjesh tanust  se ˆjroisma exwterikoÔ ginomènou dianusmˆtwn , parapèmponta se

montèla orjog¸nia aposÔnjesh tanust¸n [49℄   se montèla eÔresh bajmoÔ tanust  (bl.

Enìthta 2.3.9). To montèlo enai:

K
X
V= uj
uj
  
ujN
1 2
(4.45)

j =1

ìpou uji 2 R Ii . To prìblhma pou epiqeiretai na luje enai to:

K
min 1 jjV X
uj
uj
  
ujN jjF ; 2
uji > 0; i = 1; : : : ; N; j = 1; : : : ; K
ui 2
(4.46)
j 1 2

j =1

ìpou jjAjjF 2
enai h nìrma Frobenius 2ou bajmoÔ. H eÔresh tou pollaplasiastikoÔ epanalh-

ptikoÔ kanìna brsketai lÔnonta thn exswsh


f
uji(ii )
=0 , ìpou f enai oi ìroi sthn (4.46). O

epanalhptikì kanìna pou prokÔptei:

1 Ii 1 Ii+1 IN


PI
V Q p
i1 :::ii 1 ii+1 :::iN i1 :::l:::iN m6=i um(im )
upi(l) upi(l) PK j Q jT p (4.47)

j =1 ui(l) m= 6 i (um um )
ParathroÔme ìti to prìblhma pou epiqeire na lÔsei o Shashua enai ìmoio me to prìblhma

tou Welling [87℄, me th mình diaforˆ ston periorismì (mh arnhtikè timè ant jetik¸n tim¸n).

O epanalhptikì kanìna pou prokÔptei enai epsh parapl sio me autìn th (4.43). Gia

ton NTF algìrijmo pou protˆjhke apì tonShashua pragmatopoi jhkan peirˆmata se bˆsei

eikìnwn ( Iris image data set, Swimmer library ), sti opoe eqe prosteje jìrubo .

Oi Hazan Shashua
kai sunèqisan thn èreuna sqetikˆ me ti NTF mejìdou , aut  th forˆ

periorzonta ton arijmì twn diastˆsewn sti 3 [39℄. Protˆjhke èna algìrijmo gia NTF
qrhsimopoi¸nta thn sqetik  entropa san mètro apìstash , koin¸ thn apìklish Kullba k-
Leibler . O epanalhptikì pollaplasiastikì kanìna pou prokÔptei gia ènan tanust  3

diastˆsewn V 2 R I I I
1 2 3
enai:

up2(i2 ) up3(i3 )
PI I
2 3
i2 i3 V i i PKm
1
2 3 m m m
=1 u1(i1 ) u2(i2 ) u3(i3 )
up l up l  PI I p p : (4.48)

i2 i3 u2(i2 ) u3(i3 )
1( ) 1( ) 2 3

O kanìna dnetai antstoiqa gia ta up l


2( )
kai up l .
3( )
Oi algìrijmoi sto [39℄ sugkrnontai me

tou NMF kai PCA gia kwdikopohsh eikìnwn.

To 2006 oi Heiler kai S hnorr prìteinan ma genkeush tou algorjmou pou eqe protenei o

Hoyer [44℄ gia NMF me periorismoÔ araiìthta [42℄. O algìrijmo protˆjhke gia tanustè

2 An kai oi perissìteroi apì tou proteinìmenou algorjmou qrhsimopoioÔn ton ìro paragontopohsh

tanust¸n, kat analoga me ton NMF, pio akrib  ìro enai h aposÔnjesh tanust¸n, pou dhl¸nei ìti èna
tanust  mpore na aposunteje se ajrosmata exwterik¸n ginomènwn dianusmˆtwn.

61
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

3 diastˆsewn (me pijan  efarmog  thn kwdikopohsh eikìnwn). To prìblhma pou tjetai enai

ìmoio me autì pou protˆjhke apì tou Welling [87℄ kai Hazan [80, 39℄:

K
min 1 jjV X
uj
uj
uj jjF ; 2
uji > 0
uji 2 1 2 3
(4.49)

j =1

me th diaforˆ ìti tjetai èna epiplèon periorismì araiìthta . O Heiler orzei thn araiìthta

ìpw thn orzei o Albright sthn (4.33) [1℄. O epiplèon periorismì enai:

smin
i 6 spar(ui) 6 smax
i : (4.50)

smin max enai pragmatiko arijmo sto diˆsthma [0; 1℄ kai dnontai apì ton
kai si
Oi parˆmetroi i
qr sth, anˆloga me thn efarmog . O algìrijmo onomˆsthke SMA sparsity maximization
(

algorithm ) kai upˆrqei se morf  yeudok¸dika sto [42℄. ApodeiknÔetai ìti o algìrijmo

sugklnei se peperasmèno qronikì diˆsthma se topikì elˆqisto. Sto [42℄ pragmatopoi jhkan

peirˆmata efarmog  tou SMA se anagn¸rish pros¸pwn.

Ton IoÔnio tou 2006 o Mpoutsdh prìteine ènan algìrijmo gia thn aposÔnjesh enì

mh arnhtikoÔ tanust  [15℄. O algìrijmo onomˆsthke PALSIR proje ted alternating least
(

squares with initialization and regularization ). To prìblhma pou jètei enai to:

K
min 1 jjV X
uj
uj
uj jjF ; 2
uji > 0:
ui 2
(4.51)
j 1 2 3

j =1

Gia thn eplush tou probl mato , qrhsimopoietai o algìrijmo ALS . Apì ta dianÔsmata uji
dhmiourgoÔntai oi pnake Ui 2 R Ii K . Se kˆje epanˆlhyh, gia kˆje i = 1; 2; 3, lÔnetai

to sÔsthma th (4.51) me th mèjodo twn elaqstwn tetrag¸nwn. Epsh protenetai h idèa

efarmog  mejìdwn arqikopohsh twn tim¸n gia megalÔterh taqÔthta kai kalÔterh sÔgkli-

sh. Protenetai h qr sh arqikopoi sewn pou upˆrqoun gia thn mèjodo PARAFAC , kaj¸

kai polugrammikè epektˆsei twn mejìdwn arqikopohsh tou NMF pou prìteine o Albright
[1℄. Parìla autˆ, den gnetai kˆpoia prospˆjeia melèth twn mejìdwn arqikopohsh . Na

shmeiwje ìti sto [15℄ den dnetai apìdeixh gia ton algìrijmo kai gia thn sÔgklis  tou.

To 2007 o Ci ho ki prìteine ènan algìrijmo pou pragmatopoie paragontopohsh mh arnh-

tik¸n tanust¸n me periorismoÔ ariìthta [21℄. O algìrijmo protenetai mìno gia tanustè

3 diastˆsewn. To proteinìmeno montèlo enai to parakˆtw:

Vk = ADk Sk + Ek ; k = 1; 2; : : : ; I 3 (4.52)

Ta Vk = V ; ;k 2 R I I enai tomè (frontal sli es) tou tanust  V 2 R I I I (jewroÔme I


: :
1 2 1 2 3
3

tomè ). O A 2 R
I R enai o pnaka bˆsh (  mexh ) pou anaparistˆ tou parˆgonte pou
1

sunjètoun ta dedomèna. O Dk 2 R
RR enai diag¸nio pnaka pou èqei sthn kÔria diag¸nio

62
4.4. PROTEINŸ
OMENES MŸ
EJODOI

tou stoiqea tou pnaka D2 R I3 R . Oi pnake Sk 2 R RI anaparistoÔn ti phgè ( 


2

Tèlo , oi pnake Ek 2 R
krufè sunist¸se ) twn dedomènwn.
I I enai oi tomè tou tanust 
1 2

E 2 R I I I 1 2 3
pou perièqei ta lˆjh   ton jìrubo, anˆloga me thn efarmog . Sto Sq ma 4.5

gnetai anaparˆstash tou montèlou mh arnhtik  paragontopohsh tanust¸n pou protˆjhke

apì ton Ci ho ki .

Sq ma 4.5: Sqhmatik  anaparˆstash tou montèlou pou protenetai apì ton Ci ho ki [21℄.

O Ci ho ki qrhsimopoie san sunart sei kìstou ti - kai -apoklsei . H -apìklish


orzetai w :

Dk (Vk jjASk ) =
( ) 1 X
(vitk
[AS ℄ v + ( 1)[AS ℄ ):
1
k it itk k it
( 1)
(4.53)

itk

Na shmeiwje ìti gia = 2, h (4.53) gnetai h apìstash tou Pearson. Gia = 0:5, h (4.53)
gnetai h apìstash tou Hellinger. Gia = 1, h (4.53) gnetai h apìstash tou Neyman.
Gia !1 gnetai h genikeumènh apìklish Kullba k-Leibler (onomˆzetai kai I -apìklish).

Qrhsimopoi¸nta ma metasqhmatismènh ekdoq  th elˆttwsh katˆ klsh ( gradient des ent ),

oi epanalhptiko kanìne pou prokÔptoun enai:

 PI
1=
i=1 air (vitk =[ASk ℄it )
1

srtk srtk  PI (4.54)

q=1 aqr
1

 PI2 I3
p=1 (vip =[ASk ℄ip )
s 1=
rp
air air  PI I (4.55)

q=1 srq
2 3

Oi periorismo araiìthta mporoÔn na epiteuqjoÔn me katˆllhlo mh grammikì metasqhmatismì

th morf  srtk (srtk ) 1+ , ìpou suntelest  araiìthta .

-apìklish mpore na jewrhje san sumplhrwmatik  apìklish me thn . H sunˆrthsh


H

kìstou tou NTF algorjmou pou protenei o Ci ho ki qrhsimopoi¸nta -apìklish enai:

 
X vitk [ ASk ℄ it
[ASk ℄it vitk
Dk ) (Vk
(
jjASk ) = vitk
( + 1)
+ [ASk ℄it + 1 + Sk jjSk jjL ; 1 (4.56)

it

63
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

ìpou A enai parˆmetroi kanonikopohsh pou elègqoun ton bajmì araiìthta twn pinˆkwn
S kai A. Sthn perptwsh pou = 1 kai Sk = 0, h (4.56) ekfrˆzei thn nìrma Frobenius
2ou bajmoÔ. 'Otan ! 0 ekfrˆzetai h genikeumènh apìklish Kullba k-Leibler. Tèlo , ìtan

! 1, h (4.56) metatrèpetai sthn apìstash Itakura-Saito. Oi pollaplasiastiko kanìne


pou prokÔptoun qrhsimopoi¸nta ti -apoklsei enai:

PI 1

srtk srtk 
[ i=1 air (vitk =[ASk ℄it )
1
Sk ℄
PI (4.57)

p=1 air [ASk ℄it


1

PI
2 I3
[ p=1 (vip =[ASk ℄it )srp
1
A ℄
air air  PI I (4.58)

p=1 [ASk ℄it srp


2 3

ìpou [x℄ = maxf; xg gia thn apofug  mhdenik¸n tim¸n.

O Ci ho ki epektenei thn èreuna stou NTF algorjmou sto [22℄, ìpou protenei enalla-

ktikoÔ algorjmou gia ti die sunart sei kìstou . Protenei algorjmou basismènou

sthn ALS teqnikè kai sthn teqnik  Alternating Interior-Point Gradient . Na shmeiwje ìti

oi algìrijmoi pou dnontai aforoÔn mìno dedomèna 3 diastˆsewn. Den dnontai apodexei

sqetikˆ me thn eÔresh twn pollaplasiastik¸n kanìnwn apì ti sunart sei kìstou . Tèlo ,

na parathrhje ìti to montèlo (4.52) afenì den mpore na genikeute gia perissìtere dia-

stˆsei , afetèrou den mpore na gnei anagwg  tou montèlou sthn perptwsh 2 diastˆsewn,

ìpou jewrhtikˆ ja katèlhge sto montèlo tou NMF .

4.5 NTF Qrhsimopoi¸nta Apoklsei Bregman

O proteinìmeno algìrijmo paragontopohsh mh arnhtik¸n tanust¸n èqei tou parakˆtw

stìqou :

 Dhmiourga genikeumènou algorjmou gia efarmog  se tanustè N diastˆsewn.

 Gia N = 2, anagwg  se algìrijmo paragontopohsh mh arnhtik¸n pinˆkwn.


 Genik  lÔsh, ètsi ¸ste na enai dunat  h efarmog  diaforetik¸n sunart sewn kìstou

(pq. Eukledia apìstash, apìklish Kullba k-Leibler k.a.).

Apì ta proteinìmena montèla sthn bibliografa, protim jhke autì twn Shashua kai H-
azan [80, 39℄. To montèlo pou eqe protaje apì ton Heiler anafèretai mìno gia tanustè

3 diastˆsewn. Epsh , to montèlo tou Ci ho ki anafèretai kai autì mìno gia 3-diˆstatou

tanustè , me epiplèon prìblhma ìti den gnetai anagwg  tou probl mato gia tanustè 2h

tˆxh . O lìgo pou epilèqjhke to montèlo twn Shashua kai Hazan enai giat efarmìzetai

se N -diˆstatou tanustè kai gia N = 2 anˆgetai sto montèlo th NMF [55℄. 'Ena epiplèon

64
4.5. NTF QRHSIMOPOIŸ ISEIS BREGMAN
WNTAS APOKLŸ

pleonèkthma tou montèlou enai ìti ekfrˆzei ènan tanust  w ˆjroisma tanust¸n 1ou baj-

moÔ. 'Ara sqetzetai me parìmoia probl mata polugrammik¸n teqnik¸n anˆlush upoq¸rwn

ìpw to PARAFAC [17℄, to prìblhma th aposÔnjesh tanust¸n, kai to prìblhma th eÔre-

sh orjog¸nia aposÔnjesh tanust¸n [49℄. Praktikˆ, to montèlo parˆgei ma prosèggish

tou dojènto tanust  me kajorismèno bajmì. To meionèkthma th mejìdou twn Shashua kai

Hazan enai ìti sto [80℄ protˆjhke o algìrijmo qrhsimopoi¸nta san mètro mìno thn nìr-

maFrobenius . Antstoiqa, sto [39℄ protˆjhke algìrijmo qrhsimopoi¸nta san mètro thn

apìklishKullba k-Leibler , allˆ mìno sthn perptwsh pou N = 3.


Gia thn epteuxh genik  lÔsh , anexart tw th epilog  gia sunˆrthsh kìstou , qrh-

simopoi jhkan oi apoklsei Bregman , oi opoe apoteloÔn ma oikogèneia sunart sewn pou

ekfrˆzoun diaforetikˆ mètra [16℄. Qrhsimopoi jhkan gia thn eplush tou probl mato th

prosèggish mh arnhtik¸n pinˆkwn [83℄ kai sthn paroÔsa diatrib  ja qrhsimopoihjoÔn se èna

polugrammikì peribˆllon gia thn eplush tou pio polÔplokou probl mato th aposÔnjesh

tanust¸n.

4.5.1 Prìblhma

JewroÔme ton tanust  V2 R I1 I2 In . O stìqo th paragontopohsh mh arnhtik¸n

tanust¸n enai h aposÔnjesh tou V se èna ˆjroisma apì k tanustè 1ou bajmoÔ:
k
X
V= uj
uj
  
ujn
1 2
(4.59)

j =1

ìpou uji 2 R Ii .
+
DÔo probl mata NTF mporoÔn na dhmiourghjoÔn:

 k
X


(1) min D u
u


j j
ujn ; V (4.60)
ui >
j 0
j =1
1 2

 k
X


(2) min D V ; u
u


j j
ujn (4.61)
ui >
j 0
j =1
1 2

An analÔsoume to prìblhma (1) me bˆsh ta stoiqea tou tanust , qrhsimopoi¸nta thn

(2.68), parnoume:

 k
X
 2 IN
I1 IX k
X


D u
u


j
1
j
2
ujn ; V = D j
V
u1(i1 ) u2(i2 ) : : : ujN (iN ) ; i1 i2 :::iN
j
(4.62)

j =1 i1 ;i2 ;:::;iN j =1

Efarmìzonta ton genikì orismì gia apoklsei Bregman pou dnetai apì thn (2.66) sthn

(4.62):

65
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

 k
X
 2 IN
I1 IX  k
X


D u
u


j
1
j
2
ujn ; V =  u j
1(
j
i1 ) u2(i2 )  V
ujN (iN ) ; i1 i2 :::iN
j =1 i1 ;i2 ;:::;iN j =1
2 IN
I1 IX     k
X


 Vi i :::iN
1 2
Vi i :::iN
1 2
u j
1(
j
i1 ) u2(i2 )  ujN (iN ) Vi i :::iN
1 2
(4.63)

i1 ;i2 ;:::;iN j =1
O skopì enai h eÔresh epanalhptikoÔ pollaplasiastikoÔ kanìna gia kˆje stoiqeo

tou tanust , uji l . H ( )


exswsh (4.63) mpore na efarmoste xeqwristˆ gia kˆje uji ; i =
1; : : : ; N; j = 1; : : : ; k. Mpore na deiqje ìti:

 k
X
 Ii
X
 k
X


D u
u


j
1
j
2
ujn ; V = D u
u


j
1
j
2
uji(l)

ujn ; ii =lV (4.64)

j =1 l=1 j =1
ìpou uji(l) l-ost  tim  tou dianÔsmato uji
enai h kai Vii l 2 R I I Ii
=
1 2 1 Ii+1 IN upo-

tanust  ìpou o ii -ostì dekth isoÔtai me l .

4.5.2 Bohjhtik  Sunˆrthsh gia NTF

Orzoume thn sunˆrthsh F (ui l ) w :


( )

 k
X


F (ui l ) = D
( ) u
u


j
1
j
2
uji(l)

ujn ; ii =l V (4.65)

j =1
An efarmìsoume ton orismì twn apoklsewn Bregman ìpw fanetai sthn (2.66) mpore

na deiqje ìti:

I1 Ii X
1 Ii+1 IN  k
X
  

F (ui l ) =
( )  u j
1( i1 )  uji(l)  ujN (iN )  Vi :::ii
1 1 lii+1 :::iN
i1 ;:::;ii 1 ;ii+1 ;:::;iN j =1
I1 Ii X
1 Ii+1 IN   k
X


Vi :::ii
1 1 lii+1 :::iN u j
1( i1 )  uji(l)  ujN (iN ) Vi :::ii
1 1 lii+1 :::iN (4.66)

i1 ;:::;ii 1 ;ii+1 ;:::;iN j =1


Katanaloga me thn mèjodo prosèggish mh arnhtik¸n pinˆkwn pou prìteine o Sra , h

proteinìmenh bohjhtik  sunˆrthsh gia thn F (ui l ) enai:( )

I1 Ii X
1 Ii+1 IN  k
X

uj i    uji l    ujN iN 

G(ui l ; u~ i l ) = i :::ii jii+1 :::iN 


1( 1) ( ) ( )
( ) ( )

i1 ;:::;ii 1 ;ii+1 ;:::;iN j =1


1 1
i :::ii
1 1 jii+1 :::iN
 

 Vi :::ii
1 1 lii+1 :::iN
  k
X


Vi :::ii
1 1 lii+1 :::iN u j
1( i1 )  uji(l)  ujN (iN ) Vi :::ii
1 1 lii+1 :::iN ; (4.67)

j =1

66
4.5. NTF QRHSIMOPOIŸ ISEIS BREGMAN
WNTAS APOKLŸ

ìpou to i :::ii
1 1 jii+1 :::iN orzetai w :

uj i    u~ji l    ujN iN
i :::ii jii+1 :::iN = Pk m 1( 1) ( ) ( )

m u i u ~mil    umN iN


1 1
(4.68)

=1 1( 1) ( ) ( )

4.5.3 Apìdeixh

Ja apodexoume ìti h G(ui l ; u~ i l ) enai ìntw bohjhtik  sunˆrthsh gia thn F (ui l ). Mpo-
( ) ( ) ( )

j i :::ii jii :::iN = 1. Efìson ui ii > 0, sunepˆgetai ìti


Pk j
re eÔkola na deije ìti 1 1 +1
=1 ( )

i :::ii jii :::iN > 0. Gnwrzonta ìti h sunˆrthsh () enai kurt , isqÔei h answsh tou
1 1 +1

Jensen: (Ex) > E(x).

 Gia na apodeiqje ìti G(ui l ; ui l ) = F (ui l ):


( ) ( ) ( ) apl¸ antikajistoÔme thn exswsh (4.68)
Pk
sthn (4.67) kai qrhsimopoioÔme thn idiìthta tou i :::ii
1 1 jii+1 :::iN , ìti j =1 i1 :::ii 1 jii+1 :::iN =
1.

 Gia na apodeiqje ìti G(ui l ; u~ i l ) > F (ui l ):


( ) ( ) ( ) parathretai ìti oi ìroi th bohjhtik 

sunˆrthsh G(ui l ; u~ i l ) enai ìmoioi me tou ìrou th sunˆrthsh F (ui l ), ektì tou
( ) ( ) ( )

pr¸tou ìrou. An efarmìsoume thn answsh tou Jensen ston 1o ìro th sunˆrthsh

G(ui l ; u~ i l ):
( ) ( )

I1 Ii X
1 Ii+1 IN k
X

uj i    uji l    ujN iN 

i :::ii1 1 jii+1 :::iN 


1( 1)

i :::ii
( ) ( )
>
i1 ;:::;ii 1 ;ii+1 ;:::;iN j =1 1 1 jii+1 :::iN
I1 Ii X1 Ii+1 IN
X k  

 u j
1( i1 )  uji(l)  ujN (iN ) : (4.69)

i1 ;:::;ii 1 ;ii+1 ;:::;iN j =1

To deÔtero misì th answsh (4.69) enai o pr¸to ìro th sunˆrthsh F (ui l ).( ) 'Ara,

G(ui l ; u~ i l ) > F (ui l )


( ) ( ) ( )

Efìson apodeqjhkan ta dÔo parapˆnw shmea, sunepˆgetai ìti h sunˆrthsh G(ui l ; u~ i l )


( ) ( )

enai h bohjhtik  sunˆrthsh th F (ui l ).( )

4.5.4 Elaqistopoi¸nta thn bohjhtik  sunˆrthsh

Gia thn paragwg  enì pollaplasiastikoÔ kanìna enhmèrwsh gia to uji l , h bohjhtik  su-
( )

nˆrthsh G(ui l ; u~ i l ) ja prèpei na elaqistopoihje w pro to uji l . Gia na elaqistopoihje


( ) ( ) ( )

h sunˆrthsh G(ui l ; u ~ i l ), h merik  parˆgwgo th G(ui l ; u~ i l ) me to upi l ja prèpei na


( ) ( ) ( ) ( ) ( )
p
exiswje me to 0 kai na luje w pro to u
il: ( )

67
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

I1 Ii X
1 Ii+1 IN
G

up i    upi l    upN iN 

upi l
= i :::ii
1 1 pii+1 :::iN
1( 1)

i :::ii
( ) ( )

( ) i1 ;:::;ii 1 ;ii+1 ;:::;iN 1 1 pii+1 :::iN
up1(i1 ) : : : upi 1(ii upi p
ii+1 ) : : : uN (iN )
 i :::ii
1) +1(

1 1 pii+1 :::iN
I1 Ii X
1 Ii+1 IN    

Vi :::l:::iN 
1
u1(i1 ) : : : upi
p
up
1(ii 1 ) i+1(ii+1 )
: : : upN (iN ) (4.70)

i1 ;:::;ii 1 ;ii+1 ;:::;iN

Antikajist¸nta to i :::ii
1 1 pii+1 :::iN kai ektel¸nta ti prˆxei :

I1 Ii X
1 Ii+1 IN  
G
upi l
= u1(i1 ) : : : upi
p
up
1(ii 1 ) i+1(ii+1 )
: : : upN (iN )
( ) i1 ;:::;ii 1 ;ii+1 ;:::;iN
 k
X upi(l) 
 u1(i1 ) : : : u~i(l) : : : uN (iN ) p
m m m
u~i(l)

m=1
I1 Ii X1 Ii+1 IN    

V
i1 :::l:::iN 
up1(i1 ) : : : upi up
1(ii 1 ) i+1(ii+1 )
: : : upN (iN ) (4.71)

i1 ;:::;ii 1 ;ii+1 ;:::;iN

H exswsh (4.71) prèpei na luje w pro to stoiqeo upi l


( )
jètonta
G
upi(l) =0 .

4.5.5 Eidik  perptwsh: diaqwrsimh ()

H exswsh (4.71) den mpore na luje genikˆ gia ìle ti peript¸sei apoklsewn Bregman .

An jewrhje ìti h sunˆrthsh () enai diaqwrsimh, dhlad  ìti (xy) = (x) (y) , tìte

katal goume:

I1 Ii X
1 Ii+1 IN

upi l     k
X

( )

u~pi l
 u1(i1 ) : : : upi
p
up
1(ii 1 ) i+1(ii+1 )
: : : upN (iN ) um
1(i1 )
: : : u~mi(l) : : : um
N (iN )
( ) i1 ;:::;ii 1 ;ii+1 ;:::;iN m=1
I1 Ii X
1 Ii+1 IN    

= Vi :::l:::iN 
1
u1(i1 ) : : : upi
p
1(ii 1 )
upi+1(ii+1 ) : : : upN (iN ) (4.72)

i1 ;:::;ii 1 ;ii+1 ;:::;iN

O telikì epanalhptikì kanìna pou prokÔptei gia diaqwrsime () enai:

(Vi :::l:::iN )(up i : : : upi


P
ii+1 ) : : : uN (iN ) )
upi
 p 

upi(l) u~pi(l)  1
P
1 1( 1) 1( ii 1)
P
+1(

(up i : : : upi
1( 1) 1( ii 1)
upi +1( ii+1 ) : : : uN (iN ) )
p
( km=1 um1(i1 ) : : : u~mi(l) : : : umN(iN ))
(4.73)
P PI
1 Ii 1 Ii+1 IN
ìpou to sumbolzei to pl re ˆjroisma
i1 ;:::;ii 1 ;ii+1 ;:::;iN . O pollaplasiastikì kanìna

enhmèrwsh efarmìzetai gia ta stoiqea upi l ,


( )
ìpou p = 1; : : : ; k, i = 1; : : : ; N , kai l =
1; : : : ; Ii .

68
4.6. NTF GIA SUGKEKRIMŸ ISEIS BREGMAN
ENES APOKLŸ

4.6 NTF gia Sugkekrimène Apoklsei Bregman

4.6.1 NTF me apìklish Kullba k - Leibler

Gnwrzoume ìti (x) = x log x. 'Ara, h parˆgwgo th (x) enai (x) = log x + 1. Parath-
retai ìti (xy) = log x + log y + 1, ˆra h (xy) enai diaqwrsimh. Katal goume ètsi ston
epanalhptikì kanìna enhmèrwsh gia paragontopohsh mh arnhtik¸n tanust¸n qrhsimopoi¸-

nta thn apìklish Kullba k - Leibler :

Vi1 :::l:::iN

P
(up i : : : upi ii upi p
ii+1 ) : : : uN (iN ) ) log( km=1 um  :::u~m
P m
i(l) :::uN (iN )
)
 exp
1( 1) 1( 1) +1(

upi(l) u~pi(l) P p
1(i1 )

u1(i1 ) : : : upi 1(ii 1 ) upi+1(ii+1 ) : : : upN (iN )


(4.74)

4.6.2 NTF me nìrma Frobenius deutèrou bajmoÔ

Gnwrzoume ìti (x) = x . 'Ara, h parˆgwgo th (x) enai h (x) = x. Parathretai


1
2
2

ìti (xy) = xy, ˆra h (xy) enai diaqwrsimh. O telikì epanalhptikì kanìna gia NTF
qrhsimopoi¸nta thn nìrma Frobenius deutèrou bajmoÔ mpore na prokÔyei qrhsimopoi¸nta

thn (4.73):

upi ii : : : upN iN )  Vi :::l:::iN


P
(up i : : : upi
upi(l) u~pi(l) P 1( 1) 1( ii 1) +1( +1 ) ( ) 1

: : : upN iN )  ( km umi : : : u~mil : : : umN iN )


P (4.75)
(u1(i1 ) : : : upi
p
1(ii 1 )
upi+1(ii+1 ) ( ) =1 1( 1) ( ) ( )

4.6.3 NTF me apìstash Itakura - Saito

Gnwrzoume ìti(x) = log x. 'Ara, h parˆgwgo th (x) enai h (x) = x . 1


Fanetai ìti

(xy) = x  y , ˆra h (xy) enai diaqwrsimh. O pollaplasiastikì kanìna


1 1
enhmèrwsh

qrhsimopoi¸nta thn apìstash Itakura - Saito enai:

Ii 1 Ii+1 IN up1(i1 ) :::upi


upi+1(ii+1 ) :::upN (iN )
PI
1
i1 ;:::;ii 1 ;ii+1 ;:::;iN Pkm
=1 1(i1 ) :::u
ii
1(

um
m :::um
1)

upi l u~pi l  i(l) N (iN )


~

u p :::up u p p (4.76)
( ) ( ) PI I
1 i 1 Ii+1 IN 1( i1 ) i 1(i i 1 ) i +1( ii+1 ) :::uN (iN )
i1 ;:::;ii 1 ;ii+1 ;:::;iN Vi1 :::l:::iN

4.7 Algìrijmoi NTF gia Tanustè 3 Diastˆsewn

4.7.1 NTF 3h tˆxh me apìklish Kullba k - Leibler


Vli2 i3

PI
2
i2 =1
PI
3 p p

i3 =1 (u2(i2 ) u3(i3 ) ) log( km=1 u~m P
um um )


up l u~p l  exp PI PI
1(i1 ) 2(i2 ) 3(i3 )
(4.77)

i2 =1 i3 =1 (u2(i2 ) u3(i3 ) )
1( ) 1( ) 2 3 p p

69
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

Vi1 li3

PI
1
i1 =1
PI
3 p p
 P
i3 =1 (u1(i1 ) u3(i3 ) ) log( km=1 um u~m um )


up
l u~ p
l  exp PI PI
1(i1 ) 2(i2 ) 3(i3 )
(4.78)

i1 =1 i3 =1 (u1(i1 ) u3(i3 ) )
2( ) 2( ) 1 3 p p

Vi1 i2 l

PI
1
i1 =1
PI
2 p p
 P
i2 =1 (u1(i1 ) u2(i2 ) ) log( km=1 um um u~m )


u p
l u~ p
l  exp PI PI
1(i1 ) 2(i2 ) 3(i3 )
(4.79)

i1 =1 i2 =1 (u1(i1 ) u2(i2 ) )
3( ) 3( ) 1 2 p p

4.7.2 NTF 3h tˆxh me nìrma Frobenius 2ou bajmoÔ


 2
PI
PI
i2 =1 i3 =1 (u2(i2 ) u3(i3 ) ) li2 i3
3 p p
V 

u p
u~ p

l l PI PI Pk

(4.80)
p p
i2 =1 i3 =1 (u2(i2 ) u3(i3 ) ) ( m=1 u ~m1(i1 ) um2(i2 )um3(i3 ) )
1( ) 1( ) 2 3

 PI
PI p p
i1 =1 i3 =1 (u1(i1 ) u3(i3 ) ) i1 li3
1 3
V 

up
u~ p

l l PI PI Pk

(4.81)

i1 =1 i3 =1 (u1(i1 ) u3(i3 ) ) ( m=1 u1(i1 ) u


2( ) 2( ) 1 3 p p m ~m um )
2(i2 ) 3(i3 )

 PI
21
PI
i1 =1 i2 =1 (u1(i1 ) u2(i2 ) ) i1 i2 l
p p
V 

up
u~ p

l l PI PI Pk

(4.82)
p p
i1 =1 i2 =1 (u1(i1 ) u2(i2 ) ) ( m=1 u1(i1 ) u2(i2 ) u
3( ) 3( ) 1 2 m m ~m )
3(i3 )

4.7.3 NTF 3h tˆxh me apìstash Itakura - Saito


up2(i2 ) up3(i3 )

PI
2
PI
3
i2 =1 i3 =1 Pkm
=1 u
m um um 
up l u~p l  1(i1 ) 2(i2 ) 3(i3 )
~

1( ) 1( ) PI PI u p up (4.83)
2 3 2(i2 ) 3(i3 )
i2 =1 i3 =1 Vli2 i3
up1(i1 ) up3(i3 )

PI
1
PI
3
i1 =1 i3 =1 Pkm m ~m um 
=1 u1(i1 ) u
up l 2( )
u~p l
2( )
 PI
1
PI
3
u p up
2(i2 ) 3(i3 )

1(i1 ) 3(i3 )
(4.84)

i1 =1 i3 =1 Vi1 li3
up1(i1 ) up2(i2 )

PI
1
PI
2
i1 =1 i2 =1 Pkm m m ~m 
=1 u1(i1 ) u2(i2 ) u
up l 3( )
u~p l
3( )
 PI
1
PI
2
u p up
1( i ) 2( i )
3(i3 )
(4.85)
1 2
i1 =1 i2 =1 Vi1 i2 l

4.8 Taxinìmhsh NTF me Epbleyh

O taxinomht  NTF me epbleyh apotele genkeush tou taxinomht  NMF me epbleyh pou pou

protˆjhke sto [8℄. Baszetai sth logik  ìti pragmatopoietai h fˆsh ekpadeush gia kˆje

klˆsh qwristˆ. H fˆsh elègqou pragmatopoietai probˆllonta ta dedomèna elègqou stou

pnake pou èqoun dhmiourghje katˆ thn ekpadeush. San mètro sÔgkrish twn probol¸n

gia kˆje klˆsh qrhsimopoietai to CSM . Ston NMF ginìtan probol  twn dedomènwn elègqou

stou pnake W, stou opoou eqe upoblhje orjogwniopohsh, ètsi ¸ste ta dedomèna

elègqou na probˆllontai se orjokanonik  bˆsh. 'Omw , to prìblhma th orjogwniopohsh

tanust  enai akìma anoiktì prìblhma sth bibliografa, kai analutik  lÔsh upˆrqei mìno gia

70
4.8. TAXINŸ
OMHSH NTF ME EPŸ
IBLEYH

sugkekrimène peript¸sei tanust¸n [49℄. San apotèlesma, h diadikasa orjogwniopohsh

tou tanust  pragmatopoietai se grammikì peribˆllon, qrhsimopoi¸nta ta anaptÔgmata tou

tanust .

O algìrijmo taxinìmhsh NTF me epbleyh paratjetai sth sunèqeia. Aforˆ mìno thn

perptwsh tanust  3 diastˆsewn, allˆ mpore na epektaje kai se perissìtere diastˆsei .

JewroÔme ènan tanust  3h tˆxh V 2 R I I I


1 2 3
. O arijmì I 1 enai h diˆstash twn dedomènwn

kai C o arijmì twn klˆsewn. H proteinìmenh diadikasa taxinìmhsh me NTF apeikonzetai

sto Sq ma 4.6 kai enai h akìloujh:

1. Pragmatopohse ekpadeush se kˆje klˆsh qwristˆ:

k
X
Vi = uj
uj
uj = (U
1 2 3 i
2( ) U i ) U3( ) 3 i
1( ) (4.86)

j =1
ìpou U 1( )i enai pnaka diastˆsewn I 1( )i k (ìpou I i
1( ) o arijmì twn dedomènwn ekpa-

deush gia thn klˆsh i), o U i


2( ) enai pnaka I 2 k diastˆsewn, kai o U i
3( ) enai pnaka

diastˆsewn I3 k . To sÔmbolo sumbolzei to ginìmeno Khatri-Rao gia tanustè (bl.

enìthta 4.2.4). 'Ara, to gnìmeno (U U ) 2 3 enai tanust  diastˆsewn I 2 I k


3 .

2. Metètreye thn exswsh (4.86) se exswsh anaptÔgmato tanust¸n:

Vi = (U U )i  UT i
2 3 1( )
(4.87)

ìpou Vi enai pnaka diastˆsewn II 2 3 I i


1( ) (enai o anˆstrofo tou anaptÔgmato

tanust Vi ) kai to sÔmbolo sumbolzei to ginìmeno Khatri-Rao gia pnake .


(1) 'Ara,

o (U U )i enai pnaka diastˆsewn I I  k .


2 3 2 3

3. Ektèlese aposÔnjesh QR ston pnaka bˆsh (U U )i


2 3 :

(U U )i = Qi  Ri
2 3 (4.88)

ìpou o Qi enai orjog¸nio pnaka diastˆsewn I I  k kai o Ri enai ˆnw trigwnikì 2 3

pnaka diastˆsewn k  k . Apoj keuse tou pnake Qi kai Hi , ìpou Hi = Ri  U i .


T
1( )

4. Gia th diadikasa elègqou, jewroÔme ton pnaka qarakthristik¸n Vtest , me diastˆsei


I2 I 3, pou perièqei qarakthristikˆ gia èna dedomèno elègqou. O pnaka elègqou

probˆlletai pˆnw stou pnake bˆsh gia kˆje klˆsh:

hitest = Qyi  Vtest (4.89)

ìpou Qyi enai o Moore-Penrose yeudoantstrofo tou Qi , me diastˆsei k  I I 2 3. To

diˆnusma kwdikopohsh -elègqou hitest èqei m ko k.


71
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

5. Gia kˆje klˆsh, pragmatopoietai sÔgkrish tou dianÔsmato hitest me kˆje st lh tou

pnaka Hi , qrhsimopoi¸nta to mètro omoiìthta sunhmitìnou (CSM). To diˆnusma pou


megistopoie to CSM gia ton pnaka Hi qrhsimopoietai w mètro omoiìthta tou Vtest

me thn klˆsh i:

hitestT hji  ( )

CSMi = max
j ; ;:::;k khi kkh i k
(4.90)
( )
=1 2
test j

ìpou hji
( )
enai h j -ost  st lh tou pnaka Hi . CSMi , kajorzei thn
Tèlo , to mègisto

klˆsh sthn opoa o taxinomht  ja topojet sei dedomèno elègqou Vtest :

ClassLabel = arg i max


; ;:::;
fCSMig
=1 2
(4.91)

ht CSM
Q 1
1
H 1
1

Q ht H CSM
Vt
2 2
2 2

arg
max
. . .

. . .

. . .

h t CSM
Q H
hit = Qyi  Vt

Sq ma 4.6: Diadikasa elègqou qrhsimopoi¸nta ton proteinìmeno taxinomht  NTF (ta dia-

nÔsmata ht kai Vt sumbolzoun ta htest kai Vtest , antstoiqa).

4.9 Parathr sei

O proteinìmeno algìrijmo NTF upertere ènanti twn algorjmwn pou eqan protaje sthn

bibliografa, lìgw th genikìthtˆ tou, tìso ston arijmì twn diastˆsewn pou qrhsimopoioÔ-

ntai, ìso kai sta mètra omoiìthta pou mporoÔn na qrhsimopoihjoÔn. O algìrijmo mpore

na qrhsimopoihje gia kwdikopohsh eikìnwn kai bnteo, gia taxinìmhsh dedomènwn, akìma kai

gia omadopohsh dedomènwn. Sthn paroÔsa diatrib  ìmw , ja qrhsimopoihje sto pedo th

epexergasa  qou, èna pedo sto opoo kuriarqoÔn oi monodiˆstate teqnikè . Sugkekrimè-

na, ja qrhsimopoihje gia taxinìmhsh arqewn me bˆsh to mousikì edo sto opoo an koun,

ìpw ja perigrafe sto Kefˆlaio 6. Sthn sugkekrimènh perptwsh, kˆje mousikì arqeo ja

anaparstatai apì ènan pnaka qarakthristik¸n ston qrìno. Diaisjhtikˆ, h efarmog  NTF
72
4.9. PARATHRŸ
HSEIS

algorjmwn se tètoiou pnake epexhgetai me thn ènnoia ìti kˆje arqeo  qou apoteletai

apì grammikì sunduasmì basik¸n qarakthristik¸n. Gia parˆdeigma, to fasmatogrˆfhma a-

poteletai apì grammikì sunduasmì basik¸n suqnot twn ston qrìno. 'Etsi, apoktˆtai ma

pio analutik  melèth enì s mato  qou, kˆti pou mpore na odhg sei se beltiwmèna apote-

lèsmata se sqèsh me grammikoÔ algorjmou .

73
KEFŸ
ALAIO 4. PARAGONTOPOŸ
IHSH MH ARNHTIKŸ
WN TANUSTŸ
WN

74
Kefˆlaio 5

Exagwg  Qarakthristik¸n

5.1 Eisagwg 

Se opoiod pote sÔsthma taxinìmhsh , h epilog  twn qarakthristik¸n perigraf  twn dedo-

mènwn enai krsimh gia thn apìdosh tou taxinomht . 'Opw anafèrjhke sthn enìthta 3.3,

ta qarakthristikˆ pou qrhsimopoioÔntai sun jw se peirˆmata anagn¸rish mousikoÔ edou

katatˆssontai se 3 kathgore : qarakthristikˆ qroiˆ , rujmikˆ-qronikˆ qarakthristikˆ, kai

armonikˆ qarakthristikˆ.

Sthn paroÔsa diatrib  dnetai èmfash se qarakthristikˆ perigraf  qroiˆ . Qrhsimo-

poioÔntai perigrafe pou protenontai eurèw sto pedo th genik  anagn¸rish  qou ( ge-
neral audio re ognition - GAD ) kai sto pedo th epexergasa omila ( spee h pro essing).

O Peeters perigrˆfei kˆpoia apì ta qarakthristikˆ qroiˆ pou qrhsimopoioÔntai sthn pa-

roÔsa diatrib  sto [70℄. Epsh , qrhsimopoioÔntai basikˆ qronikˆ ( temporal ) kai armonikˆ

qarakthristikˆ, gia ma plhrèsterh perigraf  twn mousik¸n kommati¸n. Epsh , gia ìsa

qarakthristikˆ uposthrzontai, qrhsimopoi jhkan oi perigrafe pou problèpontai apì to

prìtupo MPEG-7 Audio [63℄ [71℄. 'Ola ta qarakthristikˆ pou qrhsimopoi jhkan, perigrˆ-

global
fontai w kajolikˆ ( ), me thn ènnoia ìti upologzontai gia ìlo to dosmèno s ma  

parˆjuro, lambˆnonta upìyin plhrofora apì ìla ta mousikˆ ìrgana kai tou tragou-

distè pou akoÔgontai sto mousikì kommˆti tautìqrona [78, 65℄. Ta qarakthristikˆ pou

qrhsimopoi jhkan qwrzontai sti ex  5 kathgore :

1. Qarakthristikˆ enèrgeia ( energy features )

2. Fasmatikˆ qarakthristikˆ ( spe tral features )

3. Qronikˆ qarakthristikˆ ( temporal features )

4. Antilambanìmena qarakthristikˆ ( per eptual features )

75
KEFŸ
ALAIO 5. EXAGWGŸ
H QARAKTHRISTIKŸ
WN

5. Armonikˆ qarakthristikˆ ( harmoni features )

Akolouje h dom  tou kefalaou. Sthn enìthta 5.2 dnetai ma genik  perigraf  tou

protÔpou MPEG-7 Audio kai dnetai èmfash stou perigrafe pou uposthrzei. Analuti-

k  perigraf  twn qarakthristik¸n pou qrhsimopoi jhkan gia exagwg  plhrofora apì ta

mousikˆ kommˆtia dnetai sthn enìthta 5.3 anˆ kathgora qarakthristik¸n.

5.2 To Prìtupo MPEG-7

To prìblhma th paragwg  kai diˆjesh polumesikoÔ ulikoÔ anagnwrsthke apì thn epi-

trop MPEG (Moving Pi tures Experts Group) ton IoÔlio tou 1996, ìpou apofassthke h

MPEG-7
ènarxh tou protÔpou pou enai gnwstì w , epshma onomazìmeno w  Multimedia
Content Des ription Interfa e  (Diepaf  Perigraf  PolumesikoÔ UlikoÔ) [63℄. O rìlo tou

MPEG-7 enai na kajorsei ènan sugkekrimèno trìpo perigraf  diafìrwn tÔpwn polumesi-

k  plhrofora , me skopì na katast sei pio gr gorh kai apodotik  thn anaz thsh kai thn

orgˆnwsh twn polumesik¸n dedomènwn. To prìtupo MPEG-7 qwrzetai se 8 enìthte :

1. Sust mata ( Systems )

2. Gl¸ssa perigraf  orism¸n ( Des ription De nition Language )

3. Optikì ( Visual )

4. Hqhtikì ( Audio )

5. Sq ma perigraf  polumèswn ( Multimedia Des ription S hemes )

6. Logismikì anaforˆ ( Referen e Software )

7. Sumbatìthta ( Conforman e )

8. Exagwg  kai qr sh ( Extra tion and Use )

5.2.1 Eisagwg  sto MPEG-7 Audio

To prìtupo MPEG-7 Audio apoteletai apì ergalea perigraf  ( des riptors ) kai sq mata

perigraf  ( des ription s hemes ). Ta sq mata perigraf  qwrzontai se 2 kathgore :

1. Ergalea qamhloÔ epipèdou ( low-level tools )

2. Ergalea exart¸mena apì thn efarmog  ( appli ation spe i tools )

76
5.2. TO PRŸ
OTUPO MPEG-7

Ta ergalea qamhloÔ epipèdou qarakthrzontai apì to prìtupo MPEG-7 [63, 71℄ w pla-

sio ergasa perigraf  tou  qou ( Audio Des ription Framework ). 'Eqoun thn idiìthta ìti

mporoÔn na efarmostoÔn se opoiod pote hqhtikì s ma. Audio Des ription Framework


To

perilambˆnei ta ex  :

1. KlimakoÔmene seirè ( s alable series )

2. Perigrafe qamhloÔ epipèdou ( low-level des riptors - LLDs )

3. Diast mata omoiìmorfh siwp  ( uniform silen e segment )

To Audio Des ription Framework perilambˆnei ergalea qamhloÔ epipèdou ta opoa mpo-

roÔn na qrhsimopoihjoÔn gia thn kataskeu  efarmog¸n  qou uyhlìterou epipèdou. Stìqo

tou sto MPEG-7 Audio enai na jèsei ma koin  bˆsh ergalewn suggraf  kai shmasio-

loga pˆnw sthn opoa ja mporoÔn na anaptuqjoÔn efarmogè sumbatè metaxÔ tou . H

diepaf  perigraf  qamhloÔ epipèdou ( LLD interfa e ) qwrzetai se 7 fasmatikè kai qronikè

kathgore :

1. Basikˆ Basi
( ): Stigmiae kumatomorfè kai timè enèrgeia .

2. Basikˆ Fasmatikˆ ( Basi spe tral ): To fˆsma kai fasmatikˆ qarakthristikˆ, ìpw

to kèntro bˆrou tou fˆsmato ( spe tral entroid ), h diasporˆ tou fˆsmato ( spe tral
spread ), h epipedikìthta tou fˆsmato ( spe tral atness ).

3. Parˆmetroi S mato ( Signal Parameters ): Jemeli¸dh suqnìthta hmi-periodik¸n

shmˆtwn kai armonikìthta shmˆtwn.

4. Parˆmetroi Qronik  Qroiˆ (Temporal Timbral ): Logarijmikì qrìno ènarxh

log atta k time


( ) kai qronikì kèntro bˆrou ( temporal entroid ).

5. Parˆmetroi Fasmatik  Qroiˆ ( Spe tral Timbral ): Eidikˆ fasmatikˆ qarakth-

ristikˆ se q¸ro grammik  suqnìthta pou perilambˆnoun to kèntro bˆrou tou fˆ-

smato ( spe tral entroid ), kai eidikˆ qarakthristikˆ pou anafèrontai sti armonikè

idiìthte twn shmˆtwn, ìpw to armonikì fasmatikì kèntro bˆrou ( harmoni spe-
tral entroid harmoni spe tral deviation
), armonik  fasmatik  apìklish ( ), armonik 

fasmatik  diasporˆ ( harmoni spe tral spread ) kai armonik  fasmatik  diakÔmansh

(harmoni spe tral variation ).

6. Anaparastˆsei Fasmatik  Bˆsh ( Spe tral Basis Representations ): DÔo qara-

kthristikˆ pou qrhsimopoioÔntai kurw gia anagn¸rish  qou, allˆ kai w anapara-

stˆsei tou fˆsmato se lge diastˆsei gia sumpesh plhrofora .

77
KEFŸ
ALAIO 5. EXAGWGŸ
H QARAKTHRISTIKŸ
WN

7. Diˆsthma Siwp  (Silen e Segment ): To ergaleo perigraf  th siwp  ( silen e


des riptor ) qrhsimopoietai gia anqneush siwp  se èna hqhtikì kommˆti ètsi ¸ste na

mhn epexergaste peraitèrw.

5.2.2 Perigrafe QamhloÔ Epipèdou

Oi LLDs enai mia seirˆ apì aploÔ perigrafe qamhl  poluplokìthta pou enai sqediasmè-

noi gia na qrhsimopoioÔntai apì to plasio ergasa tou diast mato  qou ( AudioSegment ).

Upˆrqoun 15 edh MPEG-7 Audio LLDs :

1. AudioLLDS alarType: Enai èna afhrhmèno orismì , pou mpore na eklhfje w

upokathgora tou klimak¸menou tÔpou ( S alar ) to opoo kai kajorzei thn tim  th peri-

graf  . 'Ena qarakthristikì tou AudioLLDS alarType enai to hopsize , me exorismoÔ

tim  10 ms
.

2. AudioLLDVe torType: Enai èna afhrhmèno orismì o opoo qrhsimopoietai boh-


jhtikˆ apì ta dianusmatikˆ ergalea perigraf  tou hqhtikoÔ s mato pou upodèqo-

ntai/epistrèfoun dianÔsmata. Enai upokathgora tou tÔpou dianÔsmato ( Ve tor ).

3. AudioWaveformType: Qrhsimopoietai gia mia oikonomik  apeikìnish th kumato-

morf  tou s mato . Qarakthristikˆ tou AudioWaveformType enai ta minRange kai

maxRange pou deqnoun ta pˆnw kai kˆtw ìria tou plˆtou tou s mato , antstoiqa.

4. AudioPowerType: Perigrˆfei th qronik¸ exomalunjesa stigmiaa isqÔ tou s mato


temporally-smoothed instantaneous power
( ).

5. AudioSpe trumAttributeGrp: Orzei èna koinì sÔnolo qarakthristik¸n pou br-

skei efarmog  se polloÔ perigrafe tou fˆsmato .

6. AudioSpe rumEnvelopeType: Perigrˆfei thn peribˆllousa tou braquqrìniou fˆ-

smato tou s mato se logarijmik  klmaka.

7. AudioSpe rumCentroidType: Kajorzei to kèntro bˆrou tou logarijmikoÔ fˆ-

smato . To Spe rumCentroid orzetai w h zugismènh isqÔ tou kèntrou bˆrou th

logarijmhmènh suqnìthta .

8. AudioSpe rumSpreadType: Perigrˆfei thn diasporˆ tou fˆsmato isqÔo loga-

rijmhmènh suqnìthta . Orzetai w h mèsh tetragwnik  apìklish tou fˆsmato isqÔo

logarijmhmènh suqnìthta , me anaforˆ sto kèntro bˆrou tou.

78
5.3. QRHSIMOPOIOŸ
UMENA QARAKTHRISTIKŸ
A

9. AudioSpe rumFlatnessType: Perigrˆfei pìso eppedo (dhlad  omoiìmorfo) enai

to braquqrìnio fˆsma isqÔo enì s mato . O perigrafèa katagrˆfei thn apìklish

tou fˆsmato isqÔo tou s mato sthn suqnìthta apì èna eppedo fˆsma (èna s ma

jorÔbou).

10. AudioSpe rumBasisType: Perilambˆnei sunart sei bˆsh sti opoe probˆlletai


èna s ma poll¸n diastˆsewn gia na prokÔyei ma apeikìnish lgwn diastˆsewn. Gia

thn oligodiˆstath apeikìnish qrhsimopoietai ete ICA , ete SVD .

11. AudioSpe rumProje tionType: Enai sumplhrwmatikì tou AudioSpe rumBasisTy-


pe kai qrhsimopoietai gia thn anaparˆstash twn qarakthristik¸n lgwn diastˆsewn

(low-dimensional features ) tou fˆsmato metˆ thn probol  tou se bˆsh meiwmènh

tˆxh .

12. AudioFundamentalFrequen yType: Perigrˆfei thn jemeli¸dh suqnìthta tou h-

qhtikoÔ s mato .

13. AudioHarmoni ityType: Perigrˆfei ton bajmì armonikìthta enì hqhtikoÔ s -

mato . Dnei dÔo qarakthristikˆ, ta opoa sunduasmèna dnoun thn armonikìthta, to

Harmoni Ratio , kai upperLimitOfHarmoni ity .

14. TimbreDes riptors: Efarmìzontai se olìklhrh thn kumatomorf  kai ìqi se degmata
tou s mato . Oi perigrafe qroiˆ baszontai ston upologismì th jemeli¸dou suqnì-

Timbre-
thta tou s mato , kai sthn eÔresh twn armonik¸n koruf¸n tou fˆsmato . Oi

Des riptors LogAtta kTime Harmoni Spe tralCentroid Harmo-


enai oi akìloujoi: , ,

ni Spe tralDeviation Harmoni Spe tralSpread Harmoni Spe tralVariation Spe tra-
, , ,

lCentroid TemporalCentroid
, kai .

15. Silen eType: H siwp  anafèretai w qarakthristikì enì hqhtikoÔ diast mato , sto

opoo den upˆrqei kajìlou shmantik  hqhtik  plhrofora.

5.3 QrhsimopoioÔmena Qarakthristikˆ

5.3.1 Qarakthristikˆ enèrgeia

MPEG-7 AudioPower (AP)


Perigrˆfei thn qronik¸ exomalunjesa stigmiaa isqÔ tou s mato ( temporally-smoothed
instantaneous power ). H stigmiaa isqÔ orzetai w :

P [n℄ = js[n℄j 2
(5.1)

79
KEFŸ
ALAIO 5. EXAGWGŸ
H QARAKTHRISTIKŸ
WN

ìpou s[n℄ to arqikì s ma. H AP upologzetai w h mèsh tim  th stigmiaa isqÔo gia

ma akolouja tim¸n m kou so me to hopsize kai apojhkeÔetai se ma qronoseirˆ mèswn

tim¸n ( Mean eld of a SeriesOfS alarType ). Parèqei mia oikonomik  apeikìnish tou fˆsmato

isqÔo , parìmoia me aut  pou prosfèrei to logarijmhmèno fˆsma isqÔo .

5.3.2 Fasmatikˆ qarakthristikˆ

MPEG-7 AudioSpe trumCentroid (ASC)


Kajorzei to kèntro bˆrou tou logarijmikoÔ fˆsmato . To kèntro bˆrou fˆsmato ( Spe-
trum Centroid ) orzetai w h zugismènh isqÔ tou kèntrou bˆrou th logarijmhmènh su-

qnìthta . To eÔro twn tim¸n tou enai apì -5 e¸ log (Fs=2000)


2
, ìpou Fs enai o rujmì

deigmatolhya . H exagwg  tou ASC :


1. Upologismì twn suntelest¸n tou fˆsmato isqÔo anˆ qronikì plasio x: Px [n℄; n =
0; :::; NF F T
2
.

2. Oi suntelestè tou fˆsmato isqÔo pou enai kˆtw apì 62.5 Hz antikajist¸ntai me

ènan suntelest , me isqÔ sh me to ˆjroismˆ tou kai onomastik  suqnìthta 31.25 Hz .

3. Oi suqnìthte ìlwn twn suntelest¸n klimak¸nontai se mia klmaka oktˆba sto 1

kHz . To ASC upologzetai w :


P

ASCx = n log2 (P
fx [n℄=1000)Px[n℄
n Px [n℄
(5.2)

ìpou f [n℄ enai h suqnìthta se Hz pou antistoiqe sto n-stì degma.


Sto ASC epishmanetai pìte to fˆsma isqÔo kuriarqetai apì qamhlè   uyhlè suqnì-

thte .

MPEG-7 AudioSpe trumSpread (ASS)


Orzetai w h mèsh tetragwnik  apìklish tou fˆsmato isqÔo logarijmhmènh suqnìthta ,

me anaforˆ sto kèntro bˆrou tou. H exagwg  tou ASS èqei w ex  :

1. Upologismì twn suntelest¸n tou fˆsmato isqÔo (ìpw orzontai sto AP ) Px (n)
kai twn antstoiqwn suqnot twn.

2. Upologismì tou ASC .


3. Upologismì th ASS w :

s
P
n ((log2 (fx [n℄P
=1000) ASCx) Px [n℄)
2

ASSx =
n Px [n℄
(5.3)

80
5.3. QRHSIMOPOIOŸ
UMENA QARAKTHRISTIKŸ
A

H diasporˆ fˆsmato enai mia oikonomik  perigraf  th morf  tou fˆsmato isqÔo

pou dhl¸nei an sugkentr¸netai kontˆ sto kèntro bˆrou tou   ektenetai se ìlo to fˆsma.

Epitrèpei ton diaqwrismì anˆmesa se jìrubo   tìno (hmitonoeidè s ma ma suqnìthta ).

MPEG-7 AudioSpe trumFlatness (ASF)


Perigrˆfei pìso eppedo (dhlad  omoiìmorfo) enai to braquqrìnio fˆsma isqÔo enì s ma-

to . To ASF katagrˆfei thn apìklish tou fˆsmato isqÔo tou s mato apì èna eppedo

fˆsma (èna s ma jorÔbou). Uyhl  apìklish deqnei thn parousa tonik¸n sustatik¸n. H a-

nˆlush omoiomorfa fˆsmato upologzetai gia ènan arijmì zwn¸n suqnìthta . H exagwg 

tou ASF apoteletai apì ta ex  b mata:

1. Pragmatopoietai fasmatik  anˆlush tou s mato qrhsimopoi¸nta , me to m ko tou

parajÔrou na antistoiqe sto hopsize (30 mse ).

2. KalÔptetai èna eÔro suqnìthta metaxÔ lowEdge kai highEdge . Oi akrae suqnìthte

twn zwn¸n metasqhmatzontai se dekte twn suntelest¸n tou fˆsmato isqÔo . Gia

kˆje mpˆnta suqnot twn, to ASF orzetai w :

q
ih(b) il(b)+1 Qih(b)
n=1 Px [n℄
ASFx;b = Pih(b) (5.4)
1
ih(b) il(b)+1 n=1 Px [n℄
'Opou il(b) kai ih(b) enai ta kat¸tata kai an¸tata ˆkra th z¸nh b. An to s ma èqei

mhdenik  mèsh enèrgeia, to ASF isoÔtai me thn monˆda.

Suqnìthta apìsbesh (spe tral roll-o frequen y -SRF)


Metrˆei se pìso uyhlè suqnìthte sto fˆsma upˆrqei èna sugkekrimèno posostì (sun jw

85%-95%) th enèrgeia tou s mato . O majhmatikì tou orismì dnetai apì th sqèsh:

 h
X K
X1

K 1
SRF = arg max S [t; k℄ < T H  S [t; k℄ (5.5)
h=0
k=0 k=0
ìpou TH enai èna kat¸fli me sun jei timè 0.85-0.95 (parnei timè sthn perioq  0-1) kai

S [t; k℄ o DFT tou s[n℄ sto t-ostì plasio. Oi toniko  qoi èqoun uyhlì SRF , en¸ oi krousto
 qoi kai h omila èqoun qamhlè timè . 'Etsi to SRF metrˆei se pìso uyhlè suqnìthte sto

fˆsma upˆrqei èna sugkekrimèno tm ma th enèrgeia .

Mel-Qasmatiko Suntelestè (Mel-frequen y epstral oeÆ ients - MFCCs)


Enai fasmatikˆ qarakthristikˆ pou baszontai ston DFT tou s mato kai thn klmaka mel,

pou montelopoioÔn thn anjr¸pinh antlhyh suqnot twn. O upologismì tou enai w ex  :

81
KEFŸ
ALAIO 5. EXAGWGŸ
H QARAKTHRISTIKŸ
WN

1. Orzetai ma trˆpeza fltrwn me N trigwnikˆ fltra:

8
>
>
>
>
0 m < f [n 1℄
>
>
>
>
>
>
>
>
>
>
>
< (f [n+1℄
>
2(m f [n 1℄)
f [n 1℄)(f [n℄ f [n 1℄)
f [n 1℄ 6 m 6 f [n℄
Hn [m℄ = (5.6)
>
>
f [n℄ 6 m 6 f [n + 1℄
>
> 2(f [n+1℄ m)
>
(f [n+1℄ f [n 1℄)(f [n+1℄ f [n℄)
>
>
>
>
>
>
>
>
>
>
>
:
0 m > f [n + 1℄

H klmaka mel orzetai w :

 
F
Bmel (F ) = 1125  ln 1 +
700 : (5.7)

Oi kentrikè suqnìthte twn fltrwn, f [n℄ me n = 1; : : : ; N , enai omoiìmorfa katane-


mhmène sthn klmaka mel:

   
Nsl B (F ) Bmel (Fl )
f [n℄ = Bmel Bmel (Fl ) + n  mel h
1
;
N +1
(5.8)
Fs
ìpou Fs enai h suqnìthta deigmatolhya se Hz, Fl enai h qamhlìterh suqnìthta th
trˆpeza fltrwn se Hz, kai Fh h antstoiqh uyhlìterh suqnìthta. Tupikè timè gia

ti Fl kai Fh enai 0 Hz kai Fs =2 Hz, antstoiqa.

2. Upologzetai h logarijmik  enèrgeia sthn èxodo tou kˆje fltrou:

Nsl 1 
X
E [n℄ = ln jS [t; k℄j Hn[k℄ ; 0 < n 6 N;
2
(5.9)

t=0

ìpou S [t; k℄ enai o DFT tou s mato sk [n℄.

3. Tèlo , upologzetai o diakritì metasqhmatismì sunhmitìnou ( DCT ) twn N logarij-

mik¸n energei¸n E [n℄:


N  

MF CC [m℄ =
X1
E [n℄ os m
2n 1 ; 06 m<N
n=0
2N (5.10)

ìpou to N parnei timè apì 24 èw 40.

82
5.3. QRHSIMOPOIOŸ
UMENA QARAKTHRISTIKŸ
A

5.3.3 Qronikˆ qarakthristikˆ

Suntelestè Autosusqètish (auto- orrelation oeÆ ients - ACs)


H autosusqètish ( auto- orrelation ) anaparistˆ thn fasmatik  katanom  tou s mato allˆ

sto pedo tou qrìnou, kaj¸ enai o antstrofo metasqhmatismì Fourier tou fˆsmato

isqÔo tou s mato . Orzetai w :

N
AC (t) =
1 X
s[n℄s[n t℄; t = 0; : : : ; N 1 (5.11)
N n=t+1
Sthn paroÔsa diatrib , krat¸ntai oi 13 pr¸toi suntelestè th autosusqètish [70℄.

MPEG-7 LogAtta kTime (LAT)


Enai logˆrijmo th qronik  diˆrkeia anˆmesa sthn qronik  stigm  pou arqzei to s ma

mèqri thn qronik  stigm  pou to s ma parnei th mègisth tim  tou. Majhmatikˆ orzetai w :

LAT = log (T 1 T 0)
10
(5.12)

ìpou T 0 enai o qrìno ènarxh tou s mato kai T 1 o qrìno pou to s ma parnei th mègisth
tim  tou. O upologismì tou LAT apeikonzetai sto Sq ma 5.1.

Sq ma 5.1: Upologismì tou LAT (apì to [63℄).

MPEG-7 TemporalCentroid (TC)


To TemporalCentroid ekfrˆzei to kèntro tou s mato ston qrìno, me bˆsh thn enèrgeia tou

s mato . Orzetai w :

TC =
PN
n=1 js[n℄j  n :
2

n js[n℄j
PN (5.13)
2
=1

Qrhsimopoietai gia ton diaqwrismì kroust¸n kai suneqìmenwn  qwn [70℄.

83
KEFŸ
ALAIO 5. EXAGWGŸ
H QARAKTHRISTIKŸ
WN

Rujmì Mhdenism¸n (zero rossing Rate - ZCR)


Oi timè tou rujmoÔ mhdenism¸n ( ZCR ) qrhsimopoioÔntai eurèw ston diaqwrismì omila kai

mousik  ( sound-musi lassi ation ). Majhmatikˆ orzetai w :

N
1 X1
ZCR =
N 1 n jsign(s[n℄) sign(s[n 1℄)j
=0
(5.14)

ìpou sign() enai h sunˆrthsh pros mou, dhlad  1 gia tou jetikoÔ arijmoÔ kai -1 gia tou

arnhtikoÔ arijmoÔ .

O algìrijmo enai aplì kai gi> autìn to lìgo èqei qamhl  upologistik  poluplokìthta.

To ZCR enai èna mètro tou jorÔbou pou perièqetai se èna s ma. Genik¸ , oi armoniko  qoi
èqoun qamhlè timè ZCR, en¸ oi krousto  qoi kai h omila èqoun uyhlè timè .

5.3.4 Antilambanìmena qarakthristikˆ

Total Loudness (TL)


H Total Loudness qrhsimopoie thn klmaka Bark gia ton upologismì twn suntelest¸n. H

klmaka Bark protˆjhke apì ton Zwi ker to 1980, se ma prospˆjeia montelopohsh tou

anjr¸pinou akoustikoÔ montèlou [90℄. H metatrop  apì Hz se suqnìthte sthn klmaka

Bark orzetai w :

   
F F
BBark (F ) = 13   tan
1315:8 +3:5   tan 7518 : (5.15)

ìpou parˆmetro , ìpou sun jw = 1:0 . Sto Sq ma 5.2 apeikonzontai oi mpˆnte th

klmaka Bark se Hz .

Gia ton upologismì th T L, h klmaka Bark qwrzetai se 24 mpˆnte diou eÔrou . Gia kˆje

suqnotik  mpˆnta upologzetai h enèrgeiˆ tou DFT pou antistoiqe sthn mpˆnta, EBark (z ),
ìpou z = 1; : : : ; 24. H Spe i Loudness orzetai w :

SL(z ) = (EBark (z )) : 0 23
(5.16)

Tèlo , h T L orzetai w to ˆjroisma twn SL gia ìle ti mpˆnte :


24
X
TL = SL(z ): (5.17)

z =1

Spe i Loudness Sensation (SLS)


Protˆjhke apì ton Pampalk to 2002 kai apotele epèktash th Total Loudness [66℄. O

upologismì twn suntelest¸n baszetai sti klmake Bark [90℄ kai Sone [11℄. Paratjetai

sunoptikˆ o upologismì twn suntelest¸n SLS :

84
5.3. QRHSIMOPOIOŸ
UMENA QARAKTHRISTIKŸ
A

Sq ma 5.2: Oi mpˆnte th klmaka Bark se Hz (apì to [70℄).

1. Upologismì tou fˆsmato isqÔo tou s mato .

2. Oi suqnìthte qwrzontai se 24 mpˆnte sÔmfwna me thn klmaka Bark .

3. Efarmog  spe tral masking sti 24 mpˆnte , ìpou prokÔptei èna s ma pou perièqei ta

tm mata me th megalÔterh enèrgeia apì ta 24 s mata.

4. Metasqhmatismì twn dedomènwn se logarijmik  klmaka.

5. Apì thn logarijmik  klmaka, upologzontai ta SL.


6. Metatrop  twn SL se monˆde sone. H SL enì tìnou 1kHz sta 40dB orzetai w 1

Sone. H klmaka sone perigrˆfetai sto [11℄.

5.3.5 Armonikˆ qarakthristikˆ

MPEG-7 AudioFundamentalFrequen y (AFF)


H jemeli¸dh suqnìthta, h opoa sumbolzetai sun jw w f 0, enai h qamhlìterh suqnìthta

se ma seirˆ armonik¸n. Oi armonikè enai pollaplˆsia th jemeli¸dou suqnìthta , kˆti

pou parathretai apì to fˆsma enì s mato , ìpou h enèrgeia tou s mato sugkentr¸netai

se pollaplˆsia th jemeli¸dou suqnìthta . Orzetai w to antstrofo th periìdou tou

85
KEFŸ
ALAIO 5. EXAGWGŸ
H QARAKTHRISTIKŸ
WN

s mato . O mousikì tìno ( pit h), enai h antilambanìmenh suqnìthta se èna s ma kai den

tautzetai me thn jemeli¸dh suqnìthta.

O pio aplì trìpo ektmhsh th jemeli¸dou suqnìthta dnetai apì thn sunˆrthsh au-

tosusqètish tou s mato [75℄. H jemeli¸dh perodo (ˆra kai h jemeli¸dh suqnìthta) tou

s mato dnetai apì to diˆsthma anˆmesa sta dÔo mègista th sunˆrthsh autosusqètish .

Mpore epsh na upologiste apì ti korufè tou DFT tou s mato . Sthn paroÔsa diatrib ,

o upologismì th jemeli¸dou suqnìthta pragmatopoi jhke qrhsimopoi¸nta algìrijmo

mègisth pijanofˆneia sto fˆsma tou s mato [27, 28℄. O upologismì th jemeli¸dou

periìdou gnetai exis¸nonta thn autosusqètish me to fˆsma isqÔo :

K  

jS (t; k)j os 2K kt ;


X1
AC (t) = 2
(5.18)

k=0

ìpou K to m ko tou DFT anˆ diˆsthma. To t pou kalÔtera exis¸nei ta dÔo mèrh dnei thn
kalÔterh ektmhsh th jemeli¸dou periìdou kai katˆ sunèpeia th jemeli¸dou suqnìthta .

86
Kefˆlaio 6

Peirˆmata

6.1 Eisagwg 

Sto parìn kefˆlaio perigrˆfetai h peiramatik  diadikasa pou efarmìsthke sthn diatrib .

Sthn enìthta 6.2 parousiˆzetai h diadikasa exagwg  qarakthristik¸n apì arqea mousikoÔ

edou kai h dhmiourga tou tanust  dedomènwn. H epilog  twn pio katˆllhlwn qarakthri-

stik¸n gia taxinìmhsh dnetai sthn enìthta 6.3. Sthn enìthta 6.4 anafèrontai sunoptikˆ oi

epiprìsjetoi taxinomhtè pou qrhsimopoi jhkan gia ta peirˆmata, pèran tou proteinìmenou

taxinomht  NTF. Ta apotelèsmata twn peiramˆtwn, h sÔgkrish me peirˆmata sthn bibliogra-

fa, kai h ermhnea twn apotelesmˆtwn dnontai sthn enìthta 6.5. Tèlo , sthn enìthta 6.6

anafèrontai telikˆ sumperˆsmata gia ta peirˆmata kai pijanè mellontikè belti¸sei .

6.2 Dhmiourga Tanust  Dedomènwn

Gia ta peirˆmata sthn diatrib  qrhsimopoi jhke h bˆsh mousikoÔ edou GTZAN , pou dhmiour-

g jhke apì ton Tzanetˆkh [84℄. 'Opw èqei  dh anaferje sthn enìthta 3.2, h bˆsh GTZAN
perièqei 1000 arqea, kalÔptonta 10 mousikˆ edh: Classi al, Country, Dis o, HipHop, Jazz,
Ro k, Blues, Reggae, Pop , kai Metal . Se kˆje klˆsh an koun 100 arqea. Kˆje arqeo èqei

katˆlhxh .au kai èqei diˆrkeia perpou 30 se


. Ta arqea enai monofwnikˆ, me suqnìthta

deigmatolhya sta 22.050 Hz , me 16 bits anˆ degma. Ta arqea èqoun thn onomatologa

ClassName.00fileNumber.au, ìpou to ClassName parnei ti timè : blues, lassi al,


ountry, dis o, hiphop, jazz, metal, pop, reggae, ro k kai to fileNumber parnei

timè apì 000 èw 100. Profan¸ , o stìqo enai h orj  taxinìmhsh twn arqewn sti pa-

rapˆnw 10 kathgore (me diamersei twn arqewn se dedomèna ekpadeush kai dedomèna

elègqou).

'Oson aforˆ thn exagwg  qarakthristik¸n apì ta arqea: se kˆje arqeo qrhsimopoietai

87
KEFŸ
ALAIO 6. PEIRŸ
AMATA

èna parˆjuro diˆrkeia 1 se , qwr epikalÔyei , ˆra kˆje arqeo qwrzetai se 30 tm mata.

Anˆ parˆjuro 1 se upologzontai 100 qarakthristikˆ, ˆra to m ko tou diast mato anˆ

tm ma enai 10 mse . Gia kˆje tm ma tou 1 se , upologzontai ta qarakthristikˆ perigraf 

pou analÔjhkan sthn enìthta 5.3. Gia ìla ta qarakthristikˆ, ektì twn AC , LAT , kai

T C, upologsthkan oi statistikè ropè pr¸th kai deÔterh tˆxh , se sunduasmì me ti

ropè twn antstoiqwn diafor¸n pr¸th tˆxh . Gia ta qarakthristikˆ AC , LAT , kai TC
den upologzontai statistikè ropè kai pr¸te diaforè , allˆ upologzontai autoÔsia. To

pl jo twn tim¸n pou parnei to kˆje qarakthristikì anˆ qronikì diˆsthma parousiˆzetai

ston Pnaka 6.1.

Pnaka 6.1: Upologismì qarakthristik¸n anˆ qronikì diˆsthma.

a/a Qarakthristikì # tim¸n/diˆsthma


1 MPEG-7 AudioPower 14
2 MPEG-7 AudioFundamentalFrequen y 14
3 Total Loudness 14
4 Spe i Loudness Sensation 84
5 MPEG-7 AudioSpe trumCentroid 14
6 Spe trum Rollo Frequen y 14
7 MPEG-7 AudioSpe trumSpread 14
8 AudioSpe trumFlattness 44
9 Mel-frequen y Cepstral CoeÆ ients 24  4
10 AutoCorrelation Values 13

11 MPEG-7 Log Atta k Time 1

12 MPEG-7 Temporal Centroid 1

13 Zero Crossing Rate 14


SÔnolo 187
'Ara, odhghj kame sthn dhmiourga enì tanust  qarakthristik¸n. O tanust  , èstw V ,

èqei diastˆsei 1000  187  30 , ìpou 1000 enai h diˆstash twn arqewn, 187 h diˆstash twn

qarakthristik¸n, kai 30 h diˆstash tou qrìnou. Sto Sq ma 6.1 dnetai ma anaparˆstash

tou tanust  dedomènwn.

6.3 Epilog  Qarakthristik¸n

Gia na elattwje to mègejo th diˆstash twn qarakthristik¸n, èna katˆllhlo uposÔnolo

twn qarakthristik¸n prèpei na epilege. To idanikì uposÔnolo qarakthristik¸n ja prèpei

88
6.3. EPILOGŸ
H QARAKTHRISTIKŸ
WN

Qar/kˆ
V
t
Dedomèna

Sq ma 6.1: Anaparˆstash tou tanust  dedomènwn.

na megistopoie ton lìgo th diasporˆ entì twn klˆsewn ( inter- lass dispersion
) pro thn

diasporˆ anˆmesa sti klˆsei ( intra- lass dispersion ). O lìgo autì sumbolzetai me:

J = tr(Sw Sb );
1
(6.1)

ìpou o telest  tr() dhl¸nei to qno enì pnaka, dhlad  to ˆjroisma twn stoiqewn th

diagwnou tou. O pnaka Sw enai o pnaka diasporˆ entì twn klˆsewn ( within- lass
s atter matrix ) kai orzetai w :

C X
X
Sw = (vk i )(vk i)T ; (6.2)

i=1 vk 2Ci

ìpou C o arijmì twn klˆsewn, vk 2 Ci dianÔsmata qarakthristik¸n gia ta dedomèna pou

an koun sthn i-ost  klˆsh, kai i to diˆnusma mèsh tim  gia ta qarakthristikˆ pou an koun
sthn i-ost  klˆsh. 'Ara o Sw enai tetragwnikì pnaka me diastˆsei ìso me to mègejo twn

qarakthristik¸n. O pnaka Sb enai o pnaka diasporˆ anˆmesa sti klˆsei (between- lass

s atter matrix ) kai orzetai w :

C
X
Sb = (i )(i )T ; (6.3)

i=1

ìpou  to diˆnusma mèsh tim  gia ta qarakthristikˆ ìlwn twn klˆsewn. Oi pnake dia-

sporˆ klˆsewn qrhsimopoioÔntai sta plasia th grammik  anˆlush diaqwrismoÔ pou pro-

tˆjhke apì ton Fisher [31℄.

'Estw ìti o arqikì arijmì qarakthristik¸n enai F, en¸ o epijumhtì arijmì qara-

kthristik¸n enai F 0, ìpou F0 < F. O arijmì twn pijan¸n uposunìlwn qarakthristik¸n

89
KEFŸ
ALAIO 6. PEIRŸ
AMATA

F 0!
pou mpore na dhmiourghjoÔn enai
(F 0 F )!F ! , kajist¸nta thn exantlhtik  anaz thsh upo-

logistikˆ apagoreutik . Gia na elattwje to upologistikì kìsto th eÔresh tou katˆl-

lhlou uposunìlou qarakthristik¸n, qrhsimopoi jhke h strathgik  anaz thsh bran h and


bound (diaklˆdwsh kai oriojèthsh ). Ston algìrijmo, dhmiourgetai ma dendroeid  dom 

(F 0 F + 1) epipèdwn, ìpou se kˆje kìmbo antistoiqe èna uposÔnolo. To uyhlìtero ep-

pedo antistoiqe sto pl re sÔnolo, en¸ sto qamhlìtero eppedo kˆje kìmbo antistoiqe

se èna uposÔnolo F0 diastˆsewn. O algìrijmo bran h and bound diasqzei th dendroeid 

dom  qrhsimopoi¸nta anaz thsh pr¸ta se bˆjo me opisjoq¸rhsh ( depth rst sear h with
ba ktra king ). Perissìtere plhrofore gia ton algìrijmo upˆrqoun sto [41℄.

Sthn sugkekrimènh perptwsh, to F isoÔtai me 187, en¸ kajorsthke metˆ apì peirˆma-

ta epijumhtì F 0 na isoÔtai me 80. Pr¸ton, ta qarakthristikˆ kanonikopoi jhkan grammikˆ


se klmaka [0; 1℄. Gia thn efarmog  tou algorjmou epilog  qarakthristik¸n, apaitetai
metatrop  tou tanust  dedomènwn se anˆptugma. Apì ton tanust  V2R 1000 187 30
dhmiour-

g jhke to anˆptugma V 2R 
(2)
187 1000 30  , sto opoo h klˆsh metabˆlletai kˆje 3.000 st le

tou pnaka. Ta 20 pr¸ta epilegmèna qarakthristikˆ paratjentai ston Pnaka 6.2. Para-

throÔme ìti ta perissìtera epilegmèna qarakthristikˆ an koun stou MF CCs, kˆti pou

deqnei giat qrhsimopoioÔntai eurèw sthn bibliografa. Axzei na shmeiwje kai h parousa

tou T L kai twn suntelest¸n AC . Oi suntelestè SLS , parìlo pou enai 32 ston arijmì,

den epilègontai ìso oi MF CCs.

6.4 Epiprìsjetoi Taxinomhtè

6.4.1 Polustrwmatiko Per eptrons

Oi polustrwmatiko Per eptrons multilayer per eptrons - MLPs


( ) apoteloÔn upokathgora

twn teqnht¸n neurwnik¸n diktÔwn ( arti ial neural networks - ANNs ). ApoteloÔntai apì

eppeda upologistik¸n monˆdwn, tou neur¸ne , ìpou kˆje eppedo neur¸nwn sundèetai me to

epìmeno ( feedforward network ). Kˆje teqnhtì neur¸na èqei pollè eisìdou , ìpou dèqetai

ta dedomèna eisìdou, èstw xi ; i = 1; : : : ; N . Ta dedomèna eisìdou stajmzontai me bˆrh,

wi ; i = 1; : : : ; N . Tèlo , kˆje neur¸na èqei ma èxodo, y , h opoa parnei thn tim  y = f (u),
PN
ìpou u =
i xi wi kai f (u) sunˆrthsh, sun jw h bhmatik    h sigmoeid  sunˆrthsh.
T
=1

Oi MLPs sun jw apoteloÔntai apì 3 eppeda, to eppedo eisìdou ( input layer ), to endiˆ-

meso eppedo ( hidden layer ), kai to eppedo exìdou ( output layer). Ma anaparˆstash enì

MLP dnetai sto Sq ma 6.2. Enai dunatìn oi MLPs me èna endiˆmeso eppedo na apeikonsoun

opoiad pote suneq  sunˆrthsh ( universal approximation theorem ).

H ekpadeush twn MLPs gnetai me th metabol  twn bar¸n tou diktÔou, qrhsimopoi¸nta

ton algìrijmo antstrofh metˆdosh ( ba kpropagation ). Diaisjhtikˆ, o skopì th ekpa-

90
6.4. EPIPRŸ
OSJETOI TAXINOMHTŸ
ES

Pnaka 6.2: 20 pr¸ta epilegmèna qarakthristikˆ tou algorjmou bran h and bound .

a/a Epilegmèno Qarakthristikì


1 Mèsh tim  2ou MF CC
2 Metablhtìthta 1ou SLS

3 Mèsh tim  2ou SLS

4 Metablhtìthta 2ou MF CC

5 5o suntelest  AC

6 Metablhtìthta 1wn diafor¸n 2ou MF CC

7 Mèsh tim  ZCR

8 Metablhtìthta 1wn diafor¸n 6ou MF CC

9 Metablhtìthta ASS

10 Mèsh tim  4ou MF CC

11 Mèsh tim  ASC

12 Mèsh tim  1ou MF CC

13 Mèsh tim  3ou MF CC

14 Metablhtìthta 3ou MF CC

15 Metablhtìthta T L

16 Mèsh tim  1wn diafor¸n 1ou MF CC

17 Mèsh tim  ASC

18 1o suntelest  AC (dhl. enèrgeia s mato )

19 Mèsh tim  T L

20 3o suntelest  AC

deush enai h èxodo tou diktÔou me dedomènh esodo na tautzetai me thn epijumht  èxodo.

Dhmiourgetai ma sunˆrthsh lˆjou kai ta bˆrh twn neur¸nwn diorj¸nontai, ètsi ¸ste na

elaqistopoietai h tim  th sunˆrthsh . H diìrjwsh twn bar¸n gnetai apì thn èxodo pro

ta psw eppeda. H diìrjwsh twn bar¸n enai epanalhptik  diadikasa kai qrhsimopoie ton

kanìna: wij (t + 1) = wij (t) + awij (t 1) + Æj yj (t), ìpou t o arijmì th epanˆlhyh , wij to
bˆro apì ton neur¸na i ston neur¸na j ,  onomˆzetai parˆmetro mˆjhsh , a enai h parˆ-

metro orm  (momentum), kai Æj to lˆjo ston neur¸na j , to opoo upologzetai anˆloga

an briskìmaste sto eppedo exìdou   an briskìmaste sto endiˆmeso eppedo. O algìrijmo

ba kpropagation upˆrqei analutikˆ sto [38℄.

91
KEFŸ
ALAIO 6. PEIRŸ
AMATA

'Exodo

Eppedo Endiˆmeso Eppedo


Eisìdou Eppedo Exìdou

Sq ma 6.2: Anaparˆstash enì Multilayer Per eptron me 4 eisìdou , 5 kìmbou sto endiˆmeso

eppedo, kai 1 èxodo.

6.4.2 Mhqanè Edrawn Dianusmˆtwn

Oi mhqanè edrawn dianusmˆtwn ( support ve tor ma hines - SVMs ) apoteloÔn èna sÔnolo

mejìdwn pou qrhsimopoioÔntai gia taxinìmhsh kai palindrìmhsh. Enai gnwsto kai w taxi-

nomhtè mègistou perijwrou ( maximum margin lassi ers ). Majanoun èna eidikì grammikì

ìrio apìfash , to upereppedo mègistou perijwrou, kˆti pou ti kajistˆ anjektikè sthn

uperprosarmog  ( over- tting ).

To upereppedo mègistou perijwrou enai ekeno to opoo epitugqˆnei to mègisto diaqwri-

smì metaxÔ twn klˆsewn. Ta shmea me th mikrìterh apìstash apì to upereppedo mègistou

perijwrou onomˆzontai edraa dianÔsmata   dianÔsmata upost rixh ( support ve tors). 'A-

ra, to sÔnolo twn edrawn dianusmˆtwn kajorzei me monadikì trìpo to upereppedo mègistou

perijwrou.

Sthn apl  perptwsh, me dedomèna pou an koun se 2 klˆsei kai enai grammik¸ diaqwr-

sima, èstw ta dedomèna fxi; yig i = 1; : : : ; N


, , me ta yi na upodeiknÔoun thn klˆsh sthn opoa
an koun ta dedomèna: yi 2 f1; 1g . An upojèsoume ìti upˆrqei kˆpoio upereppedo pou na

diaqwrzei ti 2 klˆsei , to upereppedo èqei th morf : w  x + b = 0, ìpou w diˆnusma kˆjeto


sto upereppedo. b enai parˆmetro metaknhsh tou uperepipèdou, gia na mhn dièrqetai
To

apì thn arq  twn axìnwn. A oriste d h apìstash apì to upereppedo sto kontinìtero
+

shmeo th pr¸th klˆsh kai d h apìstash apì to upereppedo sto kontinìtero shmeo

th deÔterh klˆsh . To perij¸rio (margin) tou uperepipèdou orzetai w : d + d , ìpou +

92
6.4. EPIPRŸ
OSJETOI TAXINOMHTŸ
ES

isqÔei ìti d = d = 1=jjwjj.


+ To upereppedo pou enai parˆllhlo sto upereppedo mègistou

perijwrou pou enai kontˆ sta edraa dianÔsmata th 1h klˆsh enai tow  x + b = +1, en¸
to upereppedo pou enai kontˆ sta edraa dianÔsmata th 2h klˆsh enai to w  x + b = 1.
H apìstash anˆmesa sta dÔo upereppeda enai 2=jjwjj, ˆra ja prèpei na elaqistopoihje to

jjwjj. Sto Sq ma 6.3 apeikonzetai to eppedo mègistou perijwrou gia ma perptwsh grammi-
k¸ diaqwrsimwn dedomènwn, maz me ta 2 upereppeda twn klˆsewn. An upojèsoume ìti ta

dedomèna ekpadeush enai grammik¸ diaqwrsima, tìte ikanopoioÔntai oi sunj ke :

xi  w + b > +1; gia yi = +1


xi  w + b 6 1; gia yi = 1: (6.4)

H (6.4) mpore na sumptuqje se ma exswsh:

yi(xi  w + b) 1 > 0; 8i: (6.5)

Sq ma 6.3: Apeikìnish tou uperepipèdou mègistou perijwrou gia grammik¸ diaqwrsima

dedomèna. Ta shmea me diplì kÔklo enai ta edraa dianÔsmata (apì to [18℄).

'Ara, to prìblhma t¸ra enai h elaqistopohsh tou jjwjj


, me ton periorismì th (6.5):

min 1 jjwjj ; s.t. y (x  w + b) 1 > 0:


2
i i
jjwjj 2
(6.6)

93
KEFŸ
ALAIO 6. PEIRŸ
AMATA

Gia thn eplush tou probl mato , qrhsimopoioÔme pollaplasiastè Lagrange i


, , ènan gia

kˆje periorismì th (6.5). 'Etsi, prokÔptei to prìblhma elaqistopohsh th lagkrantzian  :

n n
1
LP = jjwjj 2
X
i yi (xi  w + b) +
X
ai ;
2 i=1 i=1
(6.7)

h opoa prèpei na elaqistopoihje w pro to w, me ton periorismì i > 0. To prìblhma (6.7)

mpore na metatrape sto duðkì prìblhma megistopohsh :

n n
X 1X
LD =
i=1
ai
2 i;j i j yiyj (xi  xj );
=1
(6.8)

me tou periorismoÔ i >0 kai


Pn
i=1 i yi =0 . Ta probl mata (6.7) kai (6.8) enai pro-

bl mata tetragwnikoÔ programmatismoÔ ( quadrati programming problem ) kai mporoÔn na

epilujoÔn me sqetikè mejìdou beltistopohsh [18, 23, 79℄.

O parapˆnw taxinomht  enai grammikì , èqoun ìmw protaje kai mh grammiko algì-

rijmoi oi opooi qrhsimopoioÔn mejìdou pur na ( kernel methods ), oi opoe antikajistoÔn

to eswterikì ginìmeno dianusmˆtwn sthn (6.8) me ma mh grammik  sunˆrthsh dianusmˆtwn.

Sun jei pur ne gia SVMs enai:

1. O omogen  poluwnumikì pur na : (xi  xj )d


2. O mh omogen  poluwnumikì pur na : (xi  xj + )d
3. O pur na sunart sewn aktinik  bˆsh : exp( jjxi xj jj =(2 ))
2 2

4. O sigmoeid  pur na : tanh( (xi  xj ) + )


ìpou ;  2 R kai d 2 Z.
Epsh , en¸ oi taxinomhtè SVM efarmìzontai gia 2 klˆsei , mporoÔn eÔkola na sundua-

stoÔn gia probl mata poll¸n klˆsewn. Ma apl  teqnik  ekpaideÔei C taxinomhtè , ìpou o

kˆje èna diaqwrzei thn ma klˆsh enanton ìlwn twn ˆllwn [13℄.

6.5 Peiramatikˆ Apotelèsmata

Sta peirˆmata taxinìmhsh mousikoÔ edou qrhsimopoi jhkan sunolikˆ oi parakˆtw taxi-

nomhtè : NMF, LNMF, SNMF, SVM, MLP NTF , me apìklish Kullba k-Liebler NTF , me

nìrma Frobenius NTF


, kai Itakura-Saito
me apìstash . Gia tou algorjmou NMF qrhsi-

mopoi jhke o taxinomht  pou parousiˆsthke sthn enìthta 4.3.6. Gia ton SNMF algìrijmo

pragmatopoi jhkan peirˆmata me dÔo timè th paramètrou , me timè 0.1 (taxinomht  SNMF


1) kai 0.001 (taxinomht  SNMF 2 ).

94
6.5. PEIRAMATIKŸ
A APOTELŸ
ESMATA

Pnaka 6.3: Apotelèsmata epituqhmènh taxinìmhsh gia pnake -tanustè 187 qarakthri-

Taxinomht  Diamèrish 70%-30% Diamèrish 90%-10%


stik¸n.

NMF 54.33% 59.00%

LNMF 53.00% 52.00%

SNMF 1  = 0:1 ( ) 55.11% 60.33%

SNMF 2  = 0:001
( ) 55.44% 61.00%

SVM 66.00% (1) 68.00% (1)

MLP 67.00% (2) 67.00% (2)

NTF Kullba k-Leibler 57.00% 60.00%

NTF Frobenius 61.33% (3) 67.00% (2)

NTF Itakura-Saito 48.66% 53.00%

Sthn paroÔsa efarmog , qrhsimopoi jhke SVM poll¸n klˆsewn me mh omogen  poluw-

numikì pur na 2h tˆxh , me = 1:0. Ta dedomèna tou tanust  V metatrˆphkan se pnaka


qrhsimopoi¸nta to anˆptugma V , ìpw kai sthn epilog  qarakthristik¸n, afoÔ ta SVMs
(2)

den leitourgoÔn me poludiˆstata dedomèna, parˆ mìno me dianÔsmata. Gia thn ekpadeush

qrhsimopoi jhke o algìrijmo sequential minimal optimization (SMO) [79℄. 'Oson aforˆ

tou MLPs , qrhsimopoi jhke dktuo 3 epipèdwn, me rujmì mˆjhsh so me 0.3, me momentum
0.2, kai me 500 epanal yei gia ekpadeush. O arijmì twn neur¸nwn sto endiˆmeso eppedo

enai so me ( 187  30 + 10)=2 = 2810 . Antstoiqa, ta dedomèna tou tanust  V metatrˆphkan

se pnaka qrhsimopoi¸nta to anˆptugma V (2) .

Ta peirˆmata pragmatopoi jhkan gia tou pnake /tanustè me 187 qarakthristikˆ, ka-

j¸ kai me tou pnake /tanustè me ta 80 epilegmèna qarakthristikˆ. Qrhsimopoi jhkan 2

diaforetikè diamersei dedomènwn ekpadeush kai elègqou, h pr¸th me 70%-30% kai h deÔ-

terh me 90%-10% (sth bibliografa ta peirˆmata pragmatopoioÔntai sun jw me diamersei

90%-10%). Ston Pnaka 6.3 parousiˆzontai ta apotelèsmata epituqhmènh taxinìmhsh gia

ta dedomèna twn 187 qarakthristik¸n, en¸ ston Pnaka 6.4 parousiˆzontai ta apotelèsmata

epituqhmènh taxinìmhsh gia ta dedomèna twn 80 epilegmènwn qarakthristik¸n.

'Oson aforˆ tou algorjmou NTF , h parˆmetro bajmoÔ k epilèqjhke anˆloga me ton
algìrijmo, gia megistopohsh th akrbeia twn apotelesmˆtwn. Gia thn perptwsh twn 187

qarakthristik¸n, me 70%-30% diamèrish, to k = 65. Gia diamèrish 90%-10%, enai 62, 62,

kai 68 gia tou 3 taxinomhtè me th seirˆ pou parousiˆzontai. Gia ta 80 epilegmèna qara-

kthristikˆ me 70%-30% diamèrish, enai 60, 64,kai 65, antstoiqa. Tèlo , gia thn diamèrish

90%-10% enai 62, 66, kai 68, antstoiqa.

'Opw parathretai apì ta apotelèsmata, to uyhlìtero posostì epituqhmènh taxinìmh-

95
KEFŸ
ALAIO 6. PEIRŸ
AMATA

Pnaka 6.4: Apotelèsmata epituqhmènh taxinìmhsh gia pnake -tanustè 80 epilegmènwn

qarakthristik¸n.

Taxinomht  Diamèrish 70%-30% Diamèrish 90%-10%


NMF 58.77% 62.00%

LNMF 64.11% 64.33%

SNMF 1  = 0:1 ( ) 57.44% 64.66%

SNMF 2  = 0:001
( ) 57.55% 66.66%

SVM 64.00% (3) 73.00% (2)

MLP 65.33% (1) 72.00% (3)

NTF Kullba k-Leibler 58.66% 66.00%

NTF Frobenius 64.66% (2) 75.00% (1)

NTF Itakura-Saito 49.00% 53.00%

sh sunantˆtai apì ton taxinomht  NTF me nìrma Frobenius , me 75%. To posostì uperbanei

to 61.0% tou Tzanetˆkh sto [84℄, kai to 74.9% tou Lidy [58℄, allˆ den uperbanei to 78.5%

tou Li [57℄. O taxinomht  NTF genikìtera parousiˆzetai na èqei polÔ uyhlˆ posostˆ se

ìle ti diamersei , me posostˆ pou proseggzoun ta SVMs . O taxinomht  NTF me apì-

klish Kullba k-Leibler genikˆ èqei qamhlìterh akrbeia se ìle ti peript¸sei , me posostˆ

kontˆ stou taxinomhtè NMF . Epsh , o taxinomht  NTF me apìstash Itakura-Saito dnei

ta qamhlìtera posostˆ, akìma sugkrinìmeno kai me tou NMF algorjmou . To para-

pˆnw deqnei ìti h apìstash Itakura-Saito den enai katˆllhlh gia thn taxinìmhsh genik¸n

qarakthristik¸n kai kalÔtera na qrhsimopoietai gia thn sÔgkrish fasmˆtwn omila , ìpou

protˆjhke. Ta MLPs genikìtera èqoun akrbeia parìmoia me aut n twn SVMs . Parathretai

ìti sthn perptwsh twn 80 qarakthristik¸n me diamèrish 90%-10% ta kalÔtera apotelèsmata

dnontai apì ton taxinomht  NTF me nìrma Frobenius , ta SVMs kai ta MLPs .

Genikìtera fanetai h uperoq  twn 80 epilegmènwn qarakthristik¸n ènanti twn 187 arqi-

k¸n qarakthristik¸n. Gia ton taxinomht  NTF me nìrma Frobenius , h beltwsh se sqèsh me

ta 187 qarakthristikˆ enai +8%, gia ta SVMs enai +5%, kai gia ta MLPs +5%. Epsh ,

fanetai ìti me diamèrish 90%-10% ta apotelèsmata enai saf¸ beltiwmèna se sqèsh me thn

diamèrish 70%-30%, kˆti pou enai logikì afoÔ sthn diamèrish 90%-10% dnontai perissìtera

degmata gia ekpadeush stou taxinomhtè .

Anaforikˆ me tou taxinomhtè NMF , parathretai ìti usteroÔn ènanti twn taxinomht¸n

SVM, MLP , kai NTF me nìrma Frobenius . Sthn perptwsh twn 80 epilegmènwn qarakth-

ristik¸n, me diamèrish 90%-10%, to uyhlìtero posostì to èqei o taxinomht  SNMF 2 , me

akrbeia 66.66%. Parathretai sunolikˆ ma uperoq  tou taxinomht  SNMF 2 ènanti tou

SNMF 1 . O taxinomht  LNMF sthn perptwsh twn 187 qarakthristik¸n kai gia ti 2 dia-

96
6.6. MELLONTIKŸ
ES KATEUJŸ
UNSEIS

mersei parousiˆzei ta qamhlìtera apotelèsmata se sqèsh me ìlou tou taxinomhtè NMF .

O taxinomht  NMF tèlo , parousiˆzei akrbeia qamhlìterh apì tou taxinomhtè SNMF , me

posostˆ sugkrinìmena me ton taxinomht  NTF me apìklish Kullba k-Leibler . 'Ara, to sumpè-

rasma pou prokÔptei enai ìti oi polugrammiko algìrijmoi NTF enai saf¸ pio katˆllhloi

gia taxinìmhsh dedomènwn se sqèsh me tou grammikoÔ NMF algorjmou (profan¸ sthn

perptwsh pou ta dedomèna enai poludiˆstata).

Sthn sunèqeia, exetˆzetai h statistik  shmantikìthta twn posost¸n orj  taxinìmhsh

anˆmesa ston taxinomht  NTF me nìrma Frobenius kai tou taxinomhtè SVM kai MLP .

H mèjodo pou qrhsimopoietai parousiˆzetai sto [35℄, ìpou dnetai h upìjesh ìti ta lˆjh

taxinìmhsh gia ìlou tou taxinomhtè katanèmontai me diwnumik  katanom . Mpore na

deiqje ìti h beltwsh tou taxinomht  NTF me nìrma Frobenius se sqèsh me tou taxinomhtè

SVM kai MLP den enai statistikˆ shmantik  me diˆsthma empistosÔnh 95% ( = 0:05).
Antijètw , enai statistikˆ shmantik  se sqèsh me tou taxinomhtè NMF kai tou ˆllou

2 taxinomhtè NTF . Na shmeiwje ìti h diaforˆ 3.5% tou NTF me nìrma Frobenius kai tou

one-vs-the-rest SVM pou qrhsimopohse o Li [57℄ enai mh statistikˆ shmantik .

Analutikˆ oi epidìsei twn taxinomht¸n NTF me nìrma Frobenius SVM, , kai MLP pa-

rousiˆzontai analutikˆ qrhsimopoi¸nta pnake sÔgqush ( onfusion matri es ). O pnaka

sÔgqush gia ton taxinomht  NTF me nìrma Frobenius parousiˆzetai ston Pnaka 6.5, gia

ton taxinomht  SVM ston Pnaka 6.6, kai gia ton taxinomht  MLP ston Pnaka 6.7. Oi st -

le twn pinˆkwn sÔgqush antistoiqoÔn sto mousikì edo pou katètaxe o taxinomht  , en¸

oi grammè sto pragmatikì mousikì edo . Gia ton taxinomht  NTF me nìrma Frobenius , ta

perissìtera lˆjh gnontai gia ta mousikˆ edh Pop, Reggae Ro k


, kai . Upˆrqoun epsh 3

lˆjh taxinìmhsh apì thn Hiphop sth Dis o . Na shmeiwje ìti pollè forè ta ìria metaxÔ

eid¸n enai dusdiˆkrita, ìpw sthn perptwsh Pop kai Ro k [78℄. 'Oson aforˆ ta SVMs , oi

perissìtere lanjasmène taxinom sei sumbanoun gia to Ro k , to opoo lanjasmèna kata-

tˆssetai se Country kai Reggae. Upˆrqoun epsh lˆjh apì to Reggae pro Ro k Dis o
, ,

kai Hiphop . Akìma perissìtere lˆjo taxinom sei gia to Ro k upˆrqoun ston MLP , ìpou

katatˆssetai epsh se Country kai Reggae .

6.6 Mellontikè KateujÔnsei

Sthn paroÔsa diatrib  exetˆsthke to prìblhma th anagn¸rish mousikoÔ edou . Qrhsimo-

poi jhkan dedomèna mousikoÔ edou pou èqoun efarmoste sthn bibliografa. Protˆjhke ma

nèa teqnik  gia thn eplush tou probl mato anagn¸rish mousikoÔ edou , qrhsimopoi¸nta

tanustè . Melet jhke to episthmonikì pedo th anˆlush tanust¸n kai protˆjhke èna

nèo genikì algìrijmo gia paragontopohsh mh arnhtik¸n tanust¸n qrhsimopoi¸nta san

97
KEFŸ
ALAIO 6. PEIRŸ
AMATA

Pnaka 6.5: Pnaka sÔgqush gia ton taxinomht  NMF me nìrma Frobenius
, gia 80 epileg-

mèna qarakthristikˆ me diamèrish 90%-10%.

Edo Blues Classi al Country Dis o Hiphop Jazz Metal Pop Reggae Ro k
Blues 10 0 0 0 0 0 0 0 0 0

Classi al 0 8 1 0 0 0 0 0 1 0

Country 0 0 7 0 0 2 1 0 0 0

Dis o 1 0 0 7 0 1 0 1 0 0

Hiphop 0 0 0 3 7 0 0 0 0 0

Jazz 0 0 1 0 0 9 0 0 0 0

Metal 0 0 0 0 1 0 9 0 0 0

Pop 1 0 0 0 1 0 0 6 1 1

Reggae 0 0 1 0 2 1 0 0 6 0

Ro k 0 0 0 2 0 0 0 1 1 6

Pnaka 6.6: Pnaka sÔgqush gia ton taxinomht  SVM gia 80 epilegmèna qarakthristikˆ

me diamèrish 90%-10%.

Edo Blues Classi al Country Dis o Hiphop Jazz Metal Pop Reggae Ro k
Blues 6 1 1 0 0 0 0 0 0 0

Classi al 0 8 0 1 0 0 0 0 0 1

Country 0 0 7 0 0 0 0 0 0 2

Dis o 0 0 0 8 0 0 0 0 0 0

Hiphop 0 0 0 0 5 0 0 1 1 0

Jazz 0 1 0 1 0 10 0 0 0 0

Metal 0 0 0 2 0 0 14 0 0 0

Pop 0 0 1 0 0 0 0 6 0 0

Reggae 0 0 1 2 2 0 0 0 5 2

Ro k 1 0 3 0 0 0 0 0 3 4

98
6.6. MELLONTIKŸ
ES KATEUJŸ
UNSEIS

Pnaka 6.7: Pnaka sÔgqush gia ton taxinomht  MLP gia 80 epilegmèna qarakthristikˆ

me diamèrish 90%-10%.

Edo Blues Classi al Country Dis o Hiphop Jazz Metal Pop Reggae Ro k
Blues 7 0 0 0 0 0 1 0 0 0

Classi al 0 10 0 0 0 0 0 0 0 0

Country 1 0 6 0 0 0 0 0 0 2

Dis o 0 0 0 8 0 0 0 0 0 0

Hiphop 0 0 0 0 5 0 0 1 1 0

Jazz 0 1 0 0 0 8 0 0 2 1

Metal 0 0 0 0 1 0 13 0 0 2

Pop 0 0 0 0 1 0 0 6 0 0

Reggae 0 0 0 1 3 0 0 0 7 1

Ro k 1 0 3 0 1 0 0 0 4 2
krit rio apoklsei Bregman . Epiprìsjeta, protˆjhke èna nèo taxinomht  gia thn mèjodo

mh arnhtik  paragontopohsh tanust¸n. Oi pnake qarakthristik¸n pou dhmiourg jh-

kan perieqan fasmatikˆ qarakthristikˆ, qronikˆ qarakthristikˆ, qarakthristikˆ enèrgeia ,

qarakthristikˆ mousikoÔ tìnou kai antilambanìmena qarakthristikˆ. Ta peirˆmata pragma-

topoi jhkan qrhsimopoi¸nta kai taxinomhtè neurwnik¸n diktÔwn gia sÔgkrish. Ta apote-

lèsmata deqnoun thn uperoq  tou taxinomht  paragontopohsh mh arnhtik¸n tanust¸n se

sqèsh me grammikoÔ taxinomhtè .

Mellontikˆ, mporoÔn na gnoun belti¸sei stou proteinìmenou algorjmou , sta qara-

kthristikˆ kai sta qrhsimopoioÔmena dedomèna, gia beltwsh twn apotelesmˆtwn kai kalÔterh

sÔgkrish me th bibliografa. Sugkekrimèna:

 Qr sh dedomènwn apì ˆlle bˆsei : gia parˆdeigma oi bˆsei twn diagwnism¸n MIREX ,

ìpw oi bˆsei ISMIR 2004 kai 2005. Ektì th bˆsh GTZAN , oi upìloipe bˆsei den

èqoun qrhsimopoihje gia sugkritikˆ apotelèsmata kai den sunstatai h qr sh tou .

 Qr sh nèwn qarakthristik¸n: sthn paroÔsa diplwmatik  qrhsimopoi jhkan kurw

qarakthristikˆ qroiˆ . Oi pnake qarakthristik¸n mporoÔn na emploutistoÔn kai me

qarakthristikˆ mousikoÔ tìnou kai kurw me rujmikˆ qarakthristikˆ. Basikì rujmikì

qarakthristikì pou mpore na qrhsimopoihje enai to istìgramma periodikìthta (ana-

fèretai kai w istìgramma rujmoÔ), to opoo mpore na qrhsimopoihje kai w bˆsh gia

thn anˆptuxh nèwn rujmik¸n qarakthristik¸n.

 Anˆptuxh nèwn algorjmwn NTF , basismènoi se ˆlle apoklsei Bregman .

 Beltwsh algorjmwn NTF qrhsimopoi¸nta arqikopohsh twn pinˆkwn Ui . MporoÔn

99
KEFŸ
ALAIO 6. PEIRŸ
AMATA

na qrhsimopoihjoÔn teqnikè parìmoie me autè pou protenontai apì ton Albright [1℄.

 Beltwsh taxinomht  NTF : o proteinìmeno taxinomht  qrhsimopoie apostˆsei su-

nhmitìnou, oi opoe se qronometaballìmena qarakthristikˆ mpore na odhg soun se

lˆjh sthn taxinìmhsh. Sthn sugkekrimènh perptwsh bèbaia, jewr jhkan mikrˆ tm ma-

ta mousik¸n kommati¸n, me omoiogenè perieqìmeno. Gia beltwsh ìmw tou taxinomht ,

protenetai h qr sh istogrammˆtwn qarakthristik¸n. O taxinomht  ja sugkrnei thn

katanom  twn qarakthristik¸n gia ta dedomèna elègqou se sqèsh me ti katanomè twn

qarakthristik¸n ta opoa upˆrqoun apojhkeumèna stou pnake kwdikopohsh Hi . H

sÔgkrish twn katanom¸n mpore na gnetai qrhsimopoi¸nta thn apìstash Kullba k-


Leibler , to krit rio Kolmogorov-Smirnov ,   ˆlle parìmoie mejìdou sÔgkrish ka-

tanom¸n.

100
Bibliografa

[1℄ R. Albright, J. Cox, D. Duling, A. Langville, and C. D Meyer, \Algorithms, initializa-


tions, and onvergen e for the nonnegative matrix fa torization," preprint, 2006.
[2℄ C. A. Andersson and R. Bro, \The N-way toolbox for MATLAB,", Chemometri s and
Intelligent Laboratory Systems, Vol. 52, pp. 1-4, 2000.

[3℄ J.-J. Au outurier and F. Pa het, \Representing musi al genre: a state of the art," J.
New Musi Resear h, Vol. 32, No. 1, pp. 83-93, 2003.

[4℄ A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh, \Clustering with Bregman Diver-
gen es,", J. Ma hine Learning Resear h, Vol. 6, pp. 1705-1749, 2005.
[5℄ J. G. A. Barbedo and A. Lopes, \Automati genre lassi ation of musi al signals,"
EURASIP Journal on Advan es in Signal Pro essing, Vol. 2007, Arti le ID 64960, 2007.

[6℄ C. F. Be kmann and S. M. Smith, \Tensorial extensions of independent omponent


analysis for multi-subje t FMRI Analysis, Neuroimage, Vol. 25, No. 1, pp. 294-311,
Mar h 2005.
[7℄ E. Benetos, M. Kotti, C. Kotropoulos, J. J. Burred, G. Eisenberg, M. Haller, and T.
Sikora, \Comparison of subspa e analysis-based and statisti al model-based algorithms
for musi al instrument lassi ation," in Pro . 2nd Workshop On Immersive Commu-
ni ation And Broad ast Systems, O tober 2005.

[8℄ E. Benetos, M. Kotti, and C. Kotropoulos, \Applying supervised lassi ers based on
non-negative matrix fa torization to musi al instrument lassi ation," in Pro . IEEE
Int. Conf. Multimedia & Expo, July 2006.

[9℄ E. Benetos, M. Kotti, and C. Kotropoulos, \Large s ale musi al instrument identi a-
tion," in Pro . 4th Sound and Musi Computing Conferen e, July 2007.
[10℄ G. Birkho and S. Ma Lane, \A Survey of Modern Algebra ", New York: Ma millan
Publishing Co., 1977.
101
BIBLIOGRAFŸ
IA

[11℄ R. Bladon, \Modeling the judgment of vowel quality di eren es," J. A ousti al So iety
of Ameri a, Vol. 69, No. 5, pp. 1414-1422, 1981.

[12℄ A. I. Borisenko and I. E. Taparov, \Ve tor and Tensor Analysis with Appli ations ",
New York: Dover Publi ations In ., 1968.
[13℄ B.E. Boser, I.M. Guyon, and V. Vapnik, \A training algorithm for optimal margin
lassi ers," in Pro . 5th Annual Workshop Computational Learning Theory, pp. 144-
152, 1992.
[14℄ C. Bourin and P. Bondon, \EÆ ien y of high-order moment estimates,", IEEE Trans.
Signal Pro essing, Vol. 46, No. 1, January 1998.

[15℄ C. Boutsidis, E. Gallopoulos, P. Zhang, and R. Plemmons, "PALSIR: A new approa h to


nonnegative tensor fa torization", Poster presented at Workshop Algorithms for Modern
Massive Data Sets, June 2006.

[16℄ L. M. Bregman, \The relaxation method of nding the ommon points of onvex sets
and its appli ation to the solution of problems in onvex programming," USSR Compu-
tational Mathemati s and Mathemati al Physi s, Vol. 7, pp. 200-217, 1967.

[17℄ R. Bro, \Parafa : tutorial and appli ations,", Chemometri s and Intelligent Laboratory
Systems, Vol. 38, pp. 149-171, 1997.

[18℄ C. J. C. Burges, \A tutorial on support ve tor ma hines for pattern re ognition," Data
Mining and Knowledge Dis overy, Vol. 2, pp. 121-167, 1998.

[19℄ J.J. Burred and A. Ler h, \A hierar hi al approa h to automati musi al genre lassi-
ation," in Pro . 6th Int. Conf. Digital Audio E e ts (DAFx), September 2003.
[20℄ Z. Cataltepe, Y. Yaslan, and A. Sonmez, \Musi genre lassi ation using MIDI and
audio features," EURASIP Journal on Advan es in Signal Pro essing, Vol. 2007, Arti le
ID 36409, 2007.
[21℄ A. Ci ho ki, R. Zdunek, S. Choi, R. Plemmons, and S. Amari, \Non-negative tensor
fa torization using alpha and beta divergen es," in Pro . 32nd Int. Conf. A ousti s,
Spee h, and Signal Pro essing, April 2007.

[22℄ A. Ci ho ki, R. Zdunek, S. Choi, R. Plemmons, and S. Amari, \Novel multi-layer non-
negative tensor fa torization with sparsity onstraints," in Pro . 8th Int. Conf. Adaptive
and Natural Computing Algorithms, April 2007.

102
BIBLIOGRAFŸ
IA

[23℄ C. Cortes and V. Vapnik, \Support-Ve tor Networks," Ma hine Learning, Vol. 20, pp.
273-297, 1995.
[24℄ S. Deerwester, S. T. Dumais, G. W. Furnas, T. K. Landauer, and R. Harshman, \Inde-
xing by latent semanti analysis,", J. Ameri an So iety for Information S ien e, Vol.
41, No. 6, pp. 391-407, 1990.
[25℄ S. Dixon, F. Gouyon, and G. Widmer, \Towards hara terisation of musi via rhyth-
mi patterns," in Pro . Int. Conf. Musi Information Retrieval (ISMIR), pp. 509-516,
O tober 2004.
[26℄ D. Donoho, and V. Stodden, \When does non-negative matrix fa torization give a or-
re t de omposition into parts?," in S. Thrun, L. Saul, and B. S holkopf, eds., Advan es
in Neural Information Pro essing Systems 16, Cambridge: MIT Press, 2004.
[27℄ B. Doval and X. Rodet, \Estimation of fundamental frequen y of musi al sound signals,"
in Pro . IEEE Int. Conf. A ousti s, Spee h, and Signal Pro essing, Vol. 5, pp. 3657-3660,
April 1991.
[28℄ B. Doval and X. Rodet, \Fundamental frequen y estimation and tra king using maxi-
mum likelihood harmoni mat hing and HMMs," in Pro . IEEE Int. Conf. A ousti s,
Spee h, and Signal Pro essing, Vol. 1, pp. 221-224, April 1993.
[29℄ P. Drineas, R. Kannan, and M. W. Mahoney, \Fast Monte Carlo algorithms for matri es
III: Computing a ompressed approximate matrix de omposition,", SIAM J. Compu-
ting, Vol. 36, No. 1, pp.184-206, 2006.
[30℄ R. O. Duda, P. E. Hart, and D. G. Stork, \Pattern Classi ation," 2nd Edition, New
York: John Wiley & Sons, November 2000.
[31℄ R. A. Fisher, \The Use of Multiple Measurements in Taxonomi Problems," Annals of
Eugeni s, Vol. 7, pp. 179-188, 1936.
[32℄ E. Gaussier and C. Goutte, \Relation between PLSA and NMF and impli ations,"
in Pro . Annual ACM Conf. Resear h and Development in Information Retrieval, pp.
601-602, August 2005.
[33℄ G. H. Golub and C. F. Van Loan, \Matrix Computations," 3rd Ed., Baltimore MD:
Johns Hopkins University Press, 1996.
[34℄ E. Gomez, A. Klapuri, and B. Meudi , \Melody des ription and extra tion in the ontext
of musi ontent pro essing," J. New Musi Resear h, Vol. 32 No. 1, 2003.
103
BIBLIOGRAFŸ
IA

[35℄ I. Guyon, J. Makhoul, R. S hwartz, and V. Vapnik, \What size test set gives good error
rate estimates?,", IEEE Trans. Pattern Analysis and Ma hine Intelligen e, vol. 20, no.
1, pp. 52-64, January 1998.
[36℄ F. Gouyon, S. Dixon, E. Pampalk, and G. Widmer, \Evaluating rhythmi des riptors
for musi al genre lassi ation," in Pr . AES 25th Int. Conf., pp 196-204, June 2004.
[37℄ W. H. Greub, \Multilinear Algebra ", New York: Springer-Verlag, 1967.
[38℄ S. Haykin, \Neural networks: a omprehensive foundation," 2nd Edition, Upper Saddle
River NJ, USA: Prenti e Hall, 1999.
[39℄ T. Hazan, S. Polak, and A. Shashua, \Sparse image oding using a 3D non-negative
tensor fa torization," in Pro . 10th IEEE Int. Conf. Computer Vision, Vol. 1, pp. 50-57,
O tober 2005.
[40℄ N. He, J. Zhang, and S. Wang, \Combination of independent omponent analysis and
multi-way prin ipal omponent analysis for bat h pro ess monitoring," in Pro . 2004
IEEE Int. Conf. Systems, Man, and Cyberneti s, pp. 530-535, O tober 2004.
[41℄ F. van der Hedjen, R. P. W. Duin, D. de Ridder, and D. M. J. Tax, Classi ation,
Parameter Estimation and State Estimation, London UK: Wiley, 2004.
[42℄ M. Heiler and C. S hnorr, \Controlling sparseness in non-negative tensor fa torization,"
in Pro . 9th European Conf. Computer Vision, Vol. 1, pp. 56-67, May 2006.
[43℄ T. Hofmann, \Probabilisti latent semanti analysis," in Pro . Fifteenth Conf. Un er-
tainity in Artif ial Intelligen e, pp. 289-296, July 1999.
[44℄ P. O. Hoyer, \Non-negative matrix fa torization with sparsness onstraints," Journal
Ma hine Learning Resear h, Vol. 5, pp. 1457-1469, 2004.
[45℄ C. Hu, B. Zhang, S. Yan, Q. Yang, J. Yan, Z. Chen, and W. Ma, \Mining ratio rules
via prin ipal sparse non-negative matrix fa torization," in Pro . 2004 IEEE Int. Conf.
Data Mining, 2004.
[46℄ A. Hyvarinen and E. Oja, \Independent omponent analysis: algorithms and appli a-
tions," Neural Networks, Vol. 13, pp. 411-430, 2000.
[47℄ A. Kapteyn, H. Neude ker, and T. Wansbeek, \An approa h to n-mode omponents
analysis," Psy hometri a, Vol. 51, No. 2, pp. 269-275, June 1986.
[48℄ D. C. Kay, \Theory and Problems of Tensor Cal ulus ", New York: M Graw-Hill, 1988.
104
BIBLIOGRAFŸ
IA

[49℄ T. G. Kolda, \Orthogonal tensor de ompositions," SIAM J. Matrix Analysis Appli a-


tions, Vol. 23, No. 1, pp. 243-255, 2001.

[50℄ P. M. Kroonenberg and J. De Leeuw, \Prin ipal omponent analysis of three-mode data
by means of alternating least squares algorithms," Psy hometrika, Vol. 45, No.1, Mar h
1980.
[51℄ L. De Lathauwer, \Signal Pro essing Based on Multilinear Algebra ", Ph.D. thesis, K.U.
Leuven, E.E. Dept.-ESAT, Belgium, 1997.
[52℄ L. De Lathauwer, B. De Moor, and J. Vandewalle \An introdu tion to independent
omponent analysis," J. Chemometri s, Vol. 14, pp. 123-149, 2000.
[53℄ L. De Lathauwer, B. De Moor, and J. Vandewalle \A multilinear singular value de o-
mposition," SIAM J. Matrix Analysis Appli ations, Vol. 21, No. 4, pp. 1253-1278, April
2000.
[54℄ L. De Lathauwer, \Tensor de ompositions and independent omponent analysis," in
Pro . Workshop Tensor De ompositions, July 2004.
[55℄ D. D. Lee and H. S. Seung, \Algoritnms for non-negative matrix fa torization," Adva-
n es in Neural Information Pro essing Systems, Vol. 13, pp. 556-562, 2001.

[56℄ S. Z. Li, X. Hou, H. Zhang, and Q. Cheng, \Learning spatially lo alized, parts-based
representation," in Pro . IEEE Conf. Computer Vision and Pattern Re ognition, pp.
1-6, 2001.
[57℄ T. Li, M. Ogihara, and Q. Li, \A omparative study on ontent-based musi genre las-
si ation," in Pro . 26th Annual ACM Conf. Resear h and Development in Information
Retrieval, pp. 282-289, July-August 2003.

[58℄ T. Lidy and A. Rauber, \Evaluation of feature extra tors and psy ho-a ousti tran-
sformations for musi genre lassi ation," in Pro . 6th Int. Conf. Musi Information
Retrieval, pp. 34-41, September 2005.

[59℄ L. Lim, \Optimal solutions to non-negative PARAFAC/multilinear NMF always exist,"


in Pro . Workshop Tensor De ompositions and Appli ations, August-September 2005.
[60℄ M. Mandel and D. Ellis, \Song-level features and support ve tor ma hines for musi
lassi ation," in Pro . 6th Int. Symp. Musi Information Retrieval, pp. 594-599, Se-
ptember 2005.
105
BIBLIOGRAFŸ
IA

[61℄ C. D. M. Martin, \Tensor de ompositions workshop dis ussion notes," Ameri an Insti-
tute of Mathemati s, Palo Alto CA, July 2004.
[62℄ A. Meng, P. Ahrendt, and J. Larsen, \Improving musi genre lassi ation by short-time
feature integration," in Pro . 6th Int. Symp. Musi Information Retrieval, pp. 604-609,
September 2005.
[63℄ MPEG-7, \Information Te hnology-Multimedia Content Des ription Interfa e-Part 4:
Audio," ISO/IEC JTC1/SC29/WG11 N5525, Mar h 2003.
[64℄ F. Pa het and D. Cazaly, \A taxonomy of musi al genres," in Pro . Content-Based
Multimedia Information A ess Conf., April 2000.
[65℄ F. Pa het, J.J. Au outurier, A. La Burthe, A. Zils, and A. Beurive, \The uidado musi
browser: an end-to-end ele troni musi distribution system," Multimedia Tools and
Appli ations, Spe ial Issue on the CBMI 2003 Conf., 2004.
[66℄ E. Pampalk, A. Rauber, and D. Merkl, \Content-based organization and visualization
of musi ar hives," in Pro . 10th ACM Int. Conf. Multimedia, pp. 570-579, De ember
2002.
[67℄ E. Pampalk, A. Flexer, and G. Widmer, \Improvements of audio based musi similarity
and genre lassi ation," in Pro . 6th Int. Symp. Musi Information Retrieval, pp. 628-
633, 2005.
[68℄ A. Papoulis, \Probability, Random Variables and Sto hasti Pro esses ", New York:
M Graw-Hill, 2nd ed., 1984.
[69℄ S. W. Park and M. Savvides \Tensor fa torization by simultaneous estimation of mixing
fa tors for robust fa e re ognition and synthesis," Le ture Notes in Computer S ien e,
Vol. 4105, pp. 371-378, September 2006.
[70℄ G. Peeters, \A large set of audio features for sound des ription (similarity and lassi -
ation) in the CUIDADO proje t," CUIDADO I.S.T. Proje t Report, 2004.
[71℄ F. Pereira and R. Koenen, \Context, goals and pro edures", in Introdu tion to MPEG-
7, (B. S. Manjuntath, P. Salembier, and T. Sikora, Eds.), pp. 7-29, New York: Wiley,
2000.
[72℄ D. Perrott and R.O. Gjerdingen, \S anning the dial: An exploration of fa tors in the
identi ation of musi al style," Dept. Musi , Northwestern University, Illinois, Res.
Notes, 1999.

106
BIBLIOGRAFŸ
IA

[73℄ J. G. Proakis, C. M. Rader, F. Ling, C. L. Nikias, M. Moonen, and I. K. Proudler, \Al-


gorithms for Statisti al Signal Pro essing ", Upper Saddle River, New Jersey: Prenti e
Hall, 2002.
[74℄ A. Rauber, E. Pampalk, and D. Merkl, \Using psy ho-a ousti models and selforgani-
zing maps to reate a hierar hi al stru turing of musi by sound similarity," in Pro .
3rd Int. Conf. Musi Information Retrieval, O tober 2002.

[75℄ A. R^obel, \Fundamental frequen y estimation," Summer 2006 Le ture on Analysis,


Modeling and Transformation of Audio Sighnals, August 2006.

[76℄ F. Rousseaux and A. Bonardi, \Re on ile art and ulture on the Web: lessen the i-
mportan e of instantiation so reation an better be tion," in Pro . 1st Int. Workshop
Philosophy Informati s, April 2004.

[77℄ N. S aringella and G. Zoia, \On the modeling of time information for automati gen-
re re ognition systems in audio signals," in Pro . 6th Int. Symp. Musi Information
Retrieval, pp. 666-671, September 2005.

[78℄ N. S aringella, G. Zoia, and D. Mlynek, \Automati genre lassi ation of musi o-
ntent: a survey," IEEE Signal Pro essing Mag., Vol. 23, No. 2, pp. 133-141, Mar h
2006.
[79℄ B. S holkopf, C. J. C. Burges, and A. J. Smola, \Advan es in kernel methods: support
ve tor learning," Cambridge MA, USA: MIT Press, 1999.

[80℄ A. Shashua and T. Hazan, \Non-negative tensor fa torization with appli ations to sta-
tisti s and omputer vision," in Pro . 22nd Int. Conf. Ma hine Learning, pp. 792-799,
August 2005.
[81℄ X. Shao, C. Xu, and M. Kankanhalli, \Unsupervised lassi ation of musi al genre using
hidden Markov model," in Pro . IEEE Int. Conf. Multimedia & Expo, pp. 2023-2026,
2004.
[82℄ H. Soltau, T. S hultz, M. Westphal, and A. Waibel, \Re ognition of musi types," in
Pro . IEEE Int. Conf. A ousti s, Spee h Signal Pro essing, Vol. II, pp. 1137-1140, 1998.
[83℄ S. Sra and I. S. Dhillon, \Nonnegative matrix approximation: algorithms and appli a-
tions," Te hni al Report TR-06-27, Computer S ien es, University of Texas at Austin,
2006.
107
BIBLIOGRAFŸ
IA

[84℄ G. Tzanetakis and P. Cook, \Musi al genre lassi ation of audio signals," IEEE Trans.
Spee h and Audio Pro essing, Vol. 10, No. 5, pp. 293-302, July 2002.
[85℄ L. R. Tu ker, \The extension of fator analysis to three-dinensional matri es," in H.
Gulliksen, N. Frederiksen, Contributions to Mathemati al Psy ology, Holt, Rinehart &
Winston, N.Y., pp.109-127, 1964.
[86℄ M. A. O. Vasiles u and D. Terzopoulos, \Multilinear independent omponents analysis,"
in Pro . IEEE Conf. Computer Vision and Pattern Re ognition, Vol. 1, pp. 547-553,
June 2005.
[87℄ M. Welling and M Weber, \Positive tensor fa torization," Pattern Re ognition Letters,
Vol. 22, No. 12, pp. 1255-1261, O tober 2001.
[88℄ K. West and S. Cox, \Finding an optimal segmentation for audio genre lassi ation,"
in Pro . 6th Int. Symp. Musi Information Retrieval, pp. 680-685, September 2005.
[89℄ C. K. Yoo, D. S. Lee, P. A. Vanrolleghem, \Appli ation of multiway ICA for on-line
pro ess monitoring of a sequen ing bat h rea tor," Water Resear h, Vol. 38, pp. 1715-
1732, 2004.
[90℄ E. Zwi ker and E. Terhardt, \Analyti al expressions for riti al-band rate and riti al
bandwidth as a fun tion of frequen y," J. A ousti al So iety of Ameri a, Vol. 68, No.
5, pp. 1523-1525, June 1980.

108
Euret rio

n-ostì Bajmì Tanust , 18 Ginìmeno Tanust  me Pnaka, 12

'Olo-Orjog¸nio Tanust  , 16 Grammikì Metasqhmatismì Tanust¸n, 21

MPEG-7 AudioFundamentalFrequen y , 85
Isqurˆ Orjog¸nio Bajmì Tanust , 18
MPEG-7 AudioPower , 79
Isqurˆ Orjog¸nioi Tanustè , 15
MPEG-7 AudioSpe trumCentroid , 80

MPEG-7 AudioSpe trumFlatness , 81 Krummèna montèla Markov , 43

MPEG-7 AudioSpe trumSpread , 80


Mhqanè Edrawn Dianusmˆtwn, 92
MPEG-7 LogAtta kTime , 83
Miktì Tanust  N -ost  Tˆxh , 8
MPEG-7 TemporalCentroid , 83
Monadiao Tanust  , 10
MPEG-7 , 76
Montèla Megmato Gkaousian¸n, 43
Mel-Qasmatiko Suntelestè , 81
Mousikì edo , 31
PARAFAC , 47

Spe i Loudness Sensation , 84


Orjog¸nio Bajmì Tanust , 18

Total Loudness , 84
Pnaka , 6

Paragontopohsh mh arnhtik¸n pinˆkwn, 50


Amoibaa Orjog¸nioi Tanustè , 14
Paragontopohsh mh Arnhtik¸n Tanust¸n,
Anˆlush GrammikoÔ DiaqwrismoÔ, 43
64
Anˆptugma Tanust , 16
Pl rw Orjog¸nioi Tanustè , 15
Apì koinoÔ Rop , 22

Apì koinoÔ SusswreÔtria, 23


Polugrammik  ICA , 48

Apoklsei Bregman , 27
Polugrammik  PCA , 47

Aposuntejeimèno Tanust  , 14
Polustrwmatiko Per eptrons , 90

Prosèggish Mh Arnhtik¸n Pinˆkwn, 58

Bajmì Tanust , 18
Ropogenn tria 1h Tˆxh , 22
Bajmwtì Ginìmeno Tanust¸n, 13
Ropogenn tria 2h Tˆxh , 22
Bohjhtik  Sunˆrthsh, 58
Rujmì Mhdenism¸n, 84

Dianusmatikì Q¸ro , 5
Suntelestè Autosusqètish , 83

Suqnìthta apìsbesh , 81
Epilog  Qarakthristik¸n, 88
Sustol  Tanust , 11
Eswterikì Ginìmeno Tanust¸n, 9

Exwterikì Ginìmeno Tanust¸n, 10 Tanust  N -ost  Tˆxh , 8

109
EURETŸ
HRIO

Taxinomht  K -plhsièsterwn geitìnwn, 43


Taxinomht  NTF , 70

Uper-summetrikì Tanust  , 20

Upo-tanust  , 15

110

You might also like