• Embed Doc
  • Readcast
  • Collections
  • CommentGo Back
Download
 
Otar Verulava et al /International Journal on Computer Science and Engineering Vol.1(3), 2009, 196-198196
Prediction of the Recognition Reliability usingClustering Results
Otar Verulava, Ramaz Khurodze, Tea Todua, Otar Tavdishvili, Taliko Zhvania
Department of Informatics and Control SystemsGeorgian Technical University (GTU)Tbilisi, Georgiaverulava@gtu.ge
 Abstract
. The recognition reliability problem when realization of the recognition experiments is connected to some difficulties(some operative conditions or material expenditures) isconsidered.For estimation of recognition reliability clustering process isused. Particularly, clustering results must unambiguouslydetermine the number of clusters and their contents. The mainrequirement is validity of clustering results, which depends onthe cardinal number of set of realizations in learning sample. Theset of realizations must be representative as far as possible. Incase of satisfaction of the above-mentioned conditions it ispossible using of other procedures differing from the presented.
 Keywords: cluster, prediction, rank links, recognition reliability, Hausdorff distance
I.
 
I
NTRODUCTION
For Prediction of the recognition reliability clusteringprocess with Rank Links is used [1,2,3]. It gives us possibilityto establish the following cluster characteristics:1.
 
The number of clusters;2.
 
The number and list of realizations in each cluster;3.
 
Characteristic feature of cluster construction – a valueof cluster construction rank;4.
 
Indicator of the clusters isolation – the number of missing ranks;All four characteristics are scalars. Their values are themore valid the more are the number of realizations in learningsample of each pattern. Condition of representativeness of therealizations is a quite strong requirement and sometimes it isdifficult to satisfy it for some practical tasks. Thereforeempirical criteria is used. It implies that the more dimension of receptive field the more should be the number of realizationsin the learning samples of each pattern.Below are presented some concepts for estimation of clustering results:Definition 1. A cluster is compact if it contains only onekind of realizations, otherwise cluster is incompact.Definition 2. A pattern is compact if it is presented byonly compact clusters otherwise pattern is incompact.II.
 
P
REDICTION OF THE
R
ECOGNITION
R
ELIABILITYFOR
C
OMPACT
P
ATTERNS
 
Let’s assume that set of realizations
 X 
are clustered andthe values for all four clustering characteristics are obtained.Assume for simplicity that two elements
i
 A
and
 j
 A
from aset of patterns
 A
are taken. The task is to determine theprobabilities of correct or false recognition for these elements.Let’s assume that patterns
i
 A
and
 j
 A
are compact andtheir corresponding clusters are located as shown on Figure 1.
 g
 X 
i
 X 
  j
 A
  j
 X 
i
 A
 X 
 Figure 1.Define a minimum of the distances between therealizations of the different clusters. This distance represents aHausdorff metric between the sets. Let’s assume thatHausdorff metric is realized for realizations
ii
A X 
and
 j j
A X 
(Fig.1). Define a maximum of distances from therealizations
i
 X 
and
 j
 X 
to the realizations of their cluster,which are represented by the points-realizations
g
 X 
and
l
 X 
 (Fig.1). Let’s circumscribe the hyperspheres (circles) of radiuses
gi
 X  X 
and
l j
 X  X 
from points-realizations
i
 X 
and
 j
 X 
.Definition 3. A part of feature space, which iscircumscribed by the hypersphere of radius
gi
 X  X 
is called apattern
i
 A
influence zone on pattern
 j
 A
.
ISSN : 0975-3397
 
Otar Verulava et al /International Journal on Computer Science and Engineering Vol.1(3), 2009, 196-198197Definition 4. A part of feature space, which iscircumscribed by the hypersphere of radius
l j
 X  X 
is calleda pattern
 j
 A
influence zone on pattern
i
 A
.Take into consideration that the above-mentioneddefinitions can be used for any pairs of the set
 A
elements.Let’s consider some alternate versions of the influencezone locations:1.
 
Influence zones are disjointed.It means for patterns
i
 A
and
 j
 A
that hyperspheres withradiuses
gi
 X  X 
and
l j
 X  X 
are disjointed (Fig.1). This casecan be described by following expression:
gil jgi
 X  X  X  X  X 
<+
(1)According to Rank Links we’ll have the followinginequality:
);();();(
 jil jgi
 X  Rank  X  X  Rank  X  X  Rank 
<+
(1’)2. The influence zones are intersected (Fig. 2)
 g
 X 
i
 A
i
 X 
1
Q
2
Q
  j
 X 
 X 
 Figure 2.It’s obvious, that intersection of the influence zonesdoesn’t imply intersection of the clusters, because in this casethe condition of patterns compactness would not be satisfied.An alternate version shown on Fig. 2 can be presented bythe following inequality:
giligi
 X  X  X  X  X 
>+
(2)According to Rank Links we’ll have:
);();();(
gil jgi
 X  Rank  X  X  Rank  X  X  Rank 
>+
(2
)Prediction of the recognition reliability is based on thefollowing axioms:Axiom 1. The realizations of any pattern
 j
 A
with respectto some pattern
i
 A
can be appeared in pattern
 j
 A
influencezone only with respect to pattern
i
 A
.Axiom 2. Predictable recognition reliability is defined as aresult of division of the number of realizations belong to theinfluence zone of the given pattern on the general number of realizations of the cluster.Let’s denote the number of realizations of the pattern
i
 A
 by
i
 M 
, while the number of the realizations which has beenlocated in the influence zone of the pattern
 j
 A
by
ij
 M 
.According to axiom 2 for estimation of the recognition error
ij
P
we’ll have:
iijij
 M  M P
=
(3)where
 I  ji
,1,
=
,
.
 ji
 Based on the axiom 1, if the influence zones are notintersected, then the probability of appearance of other patternrealizations in these zones is equal to zero. Correspondingly,we’ll have
0
=
ij
 M 
. According to (3) we’ll obtain:
0
=
ij
P
 
 
1
=
+
ij
P
 Let’s use for estimation of the recognition reliability adivision of a maximal width
21
QQ
of the influence zonesintersection on the Hausdorff distance between the clusters(Fig. 2). For calculation of 
21
QQ
we’ll apply the followingexpression:
)(
2121
ji ji
QQ X  X  X QQ
+=
(4)where
,
11
Q X  X  X Q X 
 j jii
=
22
Q X  X  X  X Q
i ji j
=
.Substitute the obtained values of 
1
Q X 
i
and
 j
 X Q
2
into (4)and take into consideration that
gii
 X Q X 
=
1
and
l j j
 X Q X 
=
2
, we’ll obtain
gil j ji
 X  X  X  X  X QQ
++=
21
(5)
21
QQ
is a distance and consequently is positive. Thevalue of reliability is nonnegative scalar also. Hence we canwrite:
||
21
gil j ji
 X  X  X  X  X QQ
++=
(6)Then the recognition reliability will be:
 jigil j ji ij
 X  X  X  X  X  X  X  X  P
||
++=
(7)The same result we can obtain by (2`), written in adifferent way:
0));();(()(
<+
l jgi ji
 X  Rank  X  X  Rank  X  X  Rank 
(8)As we see from (8), the obtained difference is a negativevalue. Therefore, similar to (7), let’s use absolute value of the
ISSN : 0975-3397
of 00

Leave a Comment

You must be to leave a comment.
Submit
Characters: ...
You must be to leave a comment.
Submit
Characters: ...