Professional Documents
Culture Documents
com
Expert Systems
with Applications
Expert Systems with Applications 36 (2009) 894–908
www.elsevier.com/locate/eswa
Abstract
There have often been attempts to examine technological structure and linkage as a network. Network analysis has been mainly
employed with various centrality measures to identify core technologies in a technology network. None of the existing centrality mea-
sures, however, can successfully capture indirect relationships in a network. To address this limitation, this study proposes a novel
approach based on the analytic network process (ANP) to identification of core technologies in a technology network. Since the
ANP is capable of measuring the relative importance that captures all the indirect interactions in a network, the derived ‘‘limit centrality”
indicates the importance of a technology in terms of impacts on other technologies, taking all the direct and indirect influences into
account. The proposed approach is expected to allow technology planners to understand current technological trends and advances
by identifying core technologies based on limit centralities. Using patent citation data as proxy for interactions between technologies,
a case study on telecommunication technologies is presented to illustrate the proposed approach.
Ó 2007 Elsevier Ltd. All rights reserved.
Keywords: Analytic network process (ANP); Technology network; Core technology; Centrality; Patent citation
0957-4174/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2007.10.026
H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908 895
Network analysis has often been used in conjunction ties, the more important the entity is. Due to the highly
with patent citation analysis with the aim of grasping the skewed distribution of patent values, however, judgments
overall relationship and structure in a network. What is on importance based on simple patent counts could be
at the center of interest is to identify important or core biased to a large extent in many cases (Harhoff, Scherer,
technologies in a technology network (Shin & Park, & Vopel, 2003). It is also incapable of measuring
2007). As a quantitative measure of importance in a net- importance that mirrors influences or linkages among
work, centrality measures can be used in network analysis. entities.
Among various measures, degree centrality has been Thus, what has become the center of interest in patent
implicitly deployed as an indicator of importance of tech- analysis is citation information. Patent citation analysis is
nologies in the previous studies (Trajtenberg et al., 1997). based on the examination of citation links among different
However, it does not mirror indirect relationships despite patents (Narin, 1994). The use of citation information in
the fact that indirect citations as well as direct citations patent analysis boosts studies from various streams. One
play a crucial role in characterizing technology networks of the main research topics is to measure the values of pat-
(Wartburg et al., 2005). There are other centrality measures ents based on the number of citations of patents in subse-
that mirror indirect citations such as eigenvector centrality quent patents. It is validated by a number of evidences that
(Bonacich, 1972) and reachability out-degree (Wartburg more frequently cited patents have higher technological
et al., 2005). None of the measures, however, can success- and economic value (Breitzman & Thomas, 2002; Narin,
fully capture indirect relationships and produce meaningful Noma, & Perry, 1987; Trajtenberg, 1990). In this context,
results for identifying core technologies in a patent cita- many studies have employed the number of citations as
tion-based technology network. an indicator of patent quality (Ernst, 2003; Hirschey &
To address these limitations, this study proposes a novel Richardson, 2001; Lanjouw & Schankerman, 1999; Reit-
approach based on the analytic network process (ANP) to zig, 2004). Firm’s value can also be measured based on
identification of core technologies in a technology network. the values of patents belonging to the firm (Hall, Jaffe, &
Since the ANP is capable of measuring the relative impor- Trajtenberg, 2001). Another subject of studies with patent
tance of technologies that captures all the indirect interac- citation information is to identify similarities between tech-
tions in the technology network, the derived ‘‘limit nologies. The similarity information can be used for identi-
centrality” can be used as an implicative centrality measure fying technology overlaps with collaborative firms
characterizing a technology network and showing core (Mowery, Oxley, & Silverman, 1998), and proposing a
technologies in the network. new classification system by clustering patents (Lai &
The remainder of this paper is organized as follows. Sec- Wu, 2005).
tion 2 deals with the previous studies on patent analysis The use of patent citation information in this study is in
and centrality measures in network analysis. The underly- line with the other research stream, analyzing technological
ing methodology of the proposed approach, the ANP, is knowledge flows or technological linkages based on patent
briefly introduced in Section 3. The proposed approach is citation relationships. However, patent citation analysis
explained and illustrated with a case study in Section 4. alone cannot grasp the overall relationship and structure
The paper ends with conclusions in Section 5. among all the patents because it merely captures individual
links between two particular patents (Yoon & Park, 2004).
2. Background To address this limitation, network analysis, which will be
dealt with at the next section, has often been used in con-
2.1. Patent analysis junction with patent citation analysis to measure techno-
logical knowledge flows between entities and identify
Patents and patent statistics have long been used as tech- important or core entities. A number of studies have been
nological indicators (Grilliches, 1990). Although patents conducted at various levels, such as national level (Jaffe &
have been the representative proxy for technology as direct Trajtenberg, 1998), industry level (Han & Park, 2006), firm
output of R&D activities, there has been a ceaseless contro- level (Ham, Linden, & Appleyard, 1998), and technology
versy about the use of patent analysis since patents have class level (Shin & Park, 2007).
advantages and disadvantages like any other technological
indicator (Archibugi & Pianta, 1996). The pros and cons of 2.2. Network analysis and centrality measures
patent analysis are not explained here in detail, but can be
found in the literature by Grilliches (1990), Archibugi and In general, the interactive relationships among actors
Pianta (1996), and Ernst (2003). can be portrayed as a network composed of actors (nodes)
The most common method for early patent analysis and interactions (edges) (Gelsing, 1992). The structure of
was to simply count patents and to compare how many relations among actors and the location of actors in the
patents had been assigned to each entity, e.g. nations, network provide rich information on diverse aspects of
firms, and technological fields (Wartburg et al., 2005). an individual actor, a group of actors, and an overall
The basic idea is the more patents belong to different enti- network (Marseden & Laumann, 1984). Thus, network
896 H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908
analysis has attracted considerable interests from the social not be implicative in the technology network where cita-
and behavioral science community in recent decades, and tion frequency determines the intensity of relationships.
has also been applied and proved fruitful in a wide range In summary, it is required to develop a new centrality
of disciplines (Wasserman & Faust, 1994). A patent cita- measure that can capture indirect relationships in a
tion-based network is one of the areas where network anal- network.
ysis is effectively employed with the aim of measuring
technological knowledge flows among actors. An actor 3. Analytic network process
can be an individual patent or patents are assigned to a cor-
responding entity such as a nation or a technology class as The ANP is a generalization of the AHP (Saaty, 1996).
an actor. Then, the citation relationships among patents The AHP, also developed by Saaty (1980), is one of the most
represent interactions among actors. widely used multiple criteria decision making (MCDM)
To characterize either holistic network characteristics or methods. The AHP decomposes a problem into several lev-
individual actor’s positions in a network, various centrality els that make up a hierarchy in which each decision element
measures can be calculated. Three common measures of is supposed to be independent. The ANP extends the AHP
centrality are degree centrality, closenees centrality, and to problems with dependence and feedback. It allows for
betweenness centrality (Freeman, 1979). Among those, more complex interrelationships among decision elements
degree centrality has been implicitly deployed as an indica- by replacing a hierarchy in the AHP with a network (Meade
tor of importance of technologies in the previous studies & Sarkis, 1999). Therefore, in recent years, there has been an
(Trajtenberg et al., 1997). Degree centrality can be defined increase in the use of the ANP in a variety of problems such
as the number of ties incident upon a node. However, none as strategy selection (Wu & Lee, 2007; Yüksel & Dag devi-
of these centrality measures take into account indirect rela- ren, 2007), production-related decisions (Chung, Lee, &
tionships (Borgatti, 2005). Whereas in traditional network Pearn, 2005; Lin, Chiu, & Tsai, 2007; Mulebeke & Zheng,
theory indirect links are in general of less value than direct 2006), project selection (Cheng & Li, 2005; Lee & Kim,
links, this does not hold true in the case of patent citations 2000; Meade & Presley, 2002; Meade & Sarkis, 1999), logis-
(Wartburg et al., 2005). tics decisions (Agarwal, Shankar, & Tiwari, 2006; Gencer &
Eigenvector centrality is the one of the centrality mea- Gürpinar, 2007; Jharkharia & Shankar, 2007; Meade & Sar-
sures that has to do with indirect influence. Eigenvector kis, 1998), product design and development (Ayag& Özd-
centrality is defined as the principal eigenvector of the emir, 2007; Gun̈gor̈, 2006; Kahraman, Ertay, &
adjacency matrix defining the network (Bonacich, 1972). Büyüközkan, 2006; Karsak, Sozer, & Alptekin, 2003; Wei
Simply put, the centrality of an actor is a function of & Chang, 2007), product purchasing decision (Chang, Wu,
the centrality of actors who have relationships with the Lin, & Lin, 2007; Demirtas & Ustun, 2007), quality manage-
actor; therefore, a node that has a high eigenvector is ment (Bayazit & Karpak, 2007), and financial forecasting
one that is adjacent to nodes that are themselves high (Niemira & Saaty, 2004).
scores (Bonacich & Lloyd, 2001). Although eigenvector The process of the ANP is comprised of four major steps
centrality has become one of the standard measures of (Chung et al., 2005; Meade & Sarkis, 1999; Saaty, 1996).
network centrality, it has rarely been employed where rela-
tions among actors have different strength or intensity, (1) Network model construction. The problem is decom-
that is, valued networks, since it only considers the cen- posed into a network where nodes correspond to clus-
trality of adjacent nodes and neglects how many nodes a ters. The elements in a cluster may influence some or
node is adjacent to and how much influence a node has all the elements of any other cluster. These relation-
on adjacent nodes (Ruhnau, 2000). Also, influences on ships are represented by arcs with directions. Also,
actors that have no influence on any other actors are never
considered. Therefore, it cannot successfully capture indi-
rect relationships and produce meaningful results for tech-
nological knowledge flows in a patent citation-based
technology network.
Wartburg et al. (2005) proposed the reachability out-
degree to take into account indirect citations. The reach-
ability out-degree is defined as the probability weighted
direct Freeman out-degree times the probability weighted
direct Freeman out-degrees of the cited patents. While
the reachability out-degree can assess indirect citation
relationships, it is a proxy for specialization, not impor-
tance or impact. In addition, as what the reachability
out-degree opts for is unvalued networks where there is
no size in edges such as individual patent networks, it can- Fig. 1. Example of network in ANP and hierarchy in AHP.
H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908 897
the relationships among elements in the same cluster ment, Wij, represents a relationship between the ith
can exist and be represented by a looped arc. Fig. 1 cluster and the jth cluster. Each column of Wij is a
shows an example of the network model in the local priority vector obtained from the corresponding
ANP compared with a hierarchy in the AHP. pairwise comparison, representing the importance of
(2) Pairwise comparisons and priority vectors. Elements the elements in the ith cluster on an element in the
of each cluster are compared pairwisely with respect jth cluster. When there is no relationship between
to their impacts on an element in the cluster. In clusters, the corresponding matrix segment is a zero
addition, pairwise comparisons are made for interde- matrix.
pendency among elements outside clusters. When Then, the supermatrix is transformed into the
cluster weights are required to weight the superma- weighted supermatrix each of whose columns sums
trix at the next stage, clusters are also compared to one. This ‘column stochastic’ feature of the
pairwisely with respect to their impacts on each clus- weighted supermatrix allows convergence to occur
ter. The way of conducting pairwise comparison and in the limit supermatrix. A recommended approach
obtaining priority vectors is the same as in the AHP. to obtaining the weighted supermatrix is to deter-
The relative importance values are determined with a mine a cluster priority vector for each cluster, which
scale of 1–9, where a score of 1 indicates equal indicates relative importance of influences of other
importance between the two elements and 9 repre- clusters on each cluster. This can be done by conduct-
sents the extreme importance of one element com- ing pairwise comparisons among clusters with respect
pared to the other one. A reciprocal value is to the column cluster. The resulting priority vector is
assigned to the inverse comparison; that is, aji = then used to weight the matrix segments that fall in
1/aij, where aij denotes the importance of the ith ele- the column under the given cluster. The first entry
ment compared to the jth element. Also, aii = 1 are of the vector is multiplied by all the elements in the
preserved in the pairwise comparison matrix. Then, first matrix segment of that column, the second entry
the eigenvector method is employed to obtain local by all the elements in the second segment of the col-
priority vectors for each pairwise comparison umn and so on. Repeating this weighting procedure
matrix. for all the column clusters produces the weighted
(3) Supermatrix formation and transformation. The local supermatrix.
priority vectors are entered into the appropriate col- Finally, the weighted supermatrix is transformed
umns of a supermatrix, which is a partitioned matrix into the limit supermatrix by raising itself to
where each segment represents a relationship between powers. The reason for multiplying the weighted
two clusters. The supermatrix of a system of N clus- supermatrix is because we wish to capture the trans-
ters is denoted as the following: mission of influence along all possible paths of the
supermatrix. The entries of the weighted superma-
trix represent only the direct influence of any ele-
ment on any other element, but an element can
influence a second element indirectly through its
influence on a third element that has the direct
influence on the second element. Such one-step
indirect influences are captured by squaring the
weighted supermatrix, and two-step indirect influ-
ences are obtained from the cubic power of the
matrix, and so on. Raising the weighted supermatrix
to the power 2k + 1, where k is an arbitrarily large
number, allow convergence of the matrix, which
means the row values converge to the same value
for each column of the matrix. The resulting matrix
is called the limit supermatrix, which yields limit
priorities capturing all the indirect influences of
each element on every other element. For more
details on supermatrix characteristics and theory,
see the text by Saaty (1996)
(4) Final priorities. When the supermatrix covers the
whole network, the finial priorities of elements are
found in the corresponding columns in the limit
supermatrix. If a supermatrix only includes compo-
Ck is the kth cluster (k = 1, 2, . . ., N) which has nk nents interrelated, additional calculation should be
elements denoted as ek1, ek2, . . ., eknk. A matrix seg- made.
898 H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908
4.4. ANP
Table 1
Telecommunication technology classes 4.4.1. Network model construction
Class Title Number of mainline Basically, a network model in ANP is constructed based
subclasses on expert judgments to model an abstract decision prob-
329 Demodulators 7 lem. However, the network in the proposed approach is
331 Oscillators 37 made on the basis of citation relationships represented in
332 Modulators 7
the citation frequency matrix, as is in the case of network
340 Communications: electrical 13
341 Coded data generation or conversion 10 analysis. A cluster in the ANP network corresponds to a
342 Communications: directive radio wave 11 class, and elements in a cluster are equivalent to mainline
systems and devices subclasses in a class. In the ANP context, then, the result-
343 Communications: radio wave antennas 1 ing network model only includes alternative clusters, con-
367 Communications, electrical: acoustic wave 3
trary to the general network model in the ANP
systems and devices
370 Multiplex communications 12 comprised of a goal cluster, criteria clusters, and alternative
375 Pulse or digital communications 20 clusters. Thus, the importance of alternatives is only evalu-
379 Telephonic communications 16 ated with respect to impacts or influences on other alterna-
380 Cryptography 15 tives, not with respect to criteria or a goal, which is the
455 Telecommunications 7
same as the idea of centrality measures in network analysis.
900 H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908
Table 3
Citation frequency matrix at class level
455 380 379 375 370 367 343 342 341 340 332 331 329
455 15,487 7714 166 6804 744 20 10 344 140 1034 2837 12,884 3583
380 3303 78,954 397 1160 794 3 0 32 241 1369 114 273 176
379 1521 4514 400 208 212 36 0 26 17 279 77 200 37
375 6469 5030 29 19,587 1380 50 4 226 356 294 4466 9605 7497
370 1815 6560 58 3151 1827 35 0 54 61 310 540 1058 1229
367 59 87 0 11 0 4 0 8 55 17 3 104 15
343 94 46 0 46 0 0 128 43 0 17 16 386 18
342 844 693 3 580 33 5 18 860 4 295 191 1347 405
341 284 1132 26 770 20 12 0 12 782 227 208 1882 237
340 2145 12,084 81 644 141 0 24 300 482 5211 130 1379 314
332 1727 51 2 2630 53 5 0 33 74 15 2620 2869 509
331 10,748 566 9 7180 104 70 57 495 207 137 2045 101,992 1121
329 1495 34 0 3384 80 5 0 63 47 40 453 1373 2435
An arrow indicates the existence of patent citation rela- 4.4.2. Pairwise comparisons and priority vectors
tionships between classes or mainline subclasses. For exam- The next step deals with obtaining priority vectors.
ple, an arrow which leaves class A and enters into class B is Firstly, cluster weights are determined through comparisons
added to a network if some of the patents in class A cite at the cluster level. The basic form of measurement in the
some of the patents in class B. What this also means is class ANP is a pairwise comparison with a scale of 1–9 since sub-
B has some influences on class A; thus, subclasses of class B ject judgments have to be made on qualitative aspects.
should be pair-wisely compared with respect to impacts on However, pairwise comparisons do not have to be done
each subclass of class A. in the proposed approach. It is implicitly assumed that the
Fig. 4 shows the telecommunication technology network number of patent citations between a pair of nodes is a
for ANP including the 13 classes. Every class has influences proxy of intensity of influence. Then, the importance of
on each other, and includes a feedback loop that represents elements can be directly measured from the citation fre-
citation relationships among mainline subclasses in the quency matrix. For example, Table 3 shows that the num-
class itself. Though the network can be elaborated more ber of citations made by patents of class 455 is 3303 for
by describing citation relationships at the mainline class the patents of class 380, 1521 for the patents of class
level, it is not represented due to its complexity. 379. This can be interpreted that class 380 is about
2.17 (=3303/1.521) times more important than class 379 in is assigned to position (375, 370). In this way, the pairwise
terms of impacts on class 455. Then, the number 2.17 is comparison matrix with respect to class 455 among the 13
inserted to position (380, 379) and reciprocal value, 0.46, classes can be obtained as shown in Table 4. Then, the
Table 4
Pairwise comparison matrix with respect to class 455 and resulting priority vector
h455i 455 380 379 ... 331 329 Priority Normalization
455 1 4.69 10.18 1.44 10.36 0.3367 =15,487/45,991
380 1 2.17 0.31 2.21 0.0718 =3303/45,991
379 1 0.14 1.02 0.0331 =1521/45,991
375 0.60 4.33 0.1407 =6469/45,991
370 0.17 1.21 0.0395 =1815/45,991
367 0.01 0.04 0.0013 =59/45,991
343 0.01 0.06 0.0020 =94/45,991
342 0.08 0.56 0.0184 =844/45,991
341 0.03 0.19 0.0062 =284/45,991
340 0.20 1.43 0.0466 =2145/45,991
332 0.16 1.16 0.0376 =1727/45,991
331 1 7.19 0.2337 =10,748/45,991
329 1 0.0325 =1495/45,991
Table 5
Cluster weights
455 380 379 375 370 367 343 342 341 340 332 331 329
455 0.3367 0.0657 0.1418 0.1474 0.1381 0.0816 0.0415 0.1378 0.0568 0.1118 0.2071 0.0952 0.2039
380 0.0718 0.6721 0.3390 0.0251 0.1474 0.0122 0.0000 0.0128 0.0977 0.1481 0.0083 0.0020 0.0100
379 0.0331 0.0384 0.3416 0.0045 0.0393 0.1469 0.0000 0.0104 0.0069 0.0302 0.0056 0.0015 0.0021
375 0.1407 0.0428 0.0248 0.4244 0.2561 0.2041 0.0166 0.0905 0.1444 0.0318 0.3260 0.0710 0.4265
370 0.0395 0.0558 0.0495 0.0683 0.3391 0.1429 0.0000 0.0216 0.0247 0.0335 0.0394 0.0078 0.0699
367 0.0013 0.0007 0.0000 0.0002 0.0000 0.0163 0.0000 0.0032 0.0223 0.0018 0.0002 0.0008 0.0009
343 0.0020 0.0004 0.0000 0.0010 0.0000 0.0000 0.5311 0.0172 0.0000 0.0018 0.0012 0.0029 0.0010
342 0.0184 0.0059 0.0026 0.0126 0.0061 0.0204 0.0747 0.3446 0.0016 0.0319 0.0139 0.0100 0.0230
341 0.0062 0.0096 0.0222 0.0167 0.0037 0.0490 0.0000 0.0048 0.3171 0.0246 0.0152 0.0139 0.0135
340 0.0466 0.1029 0.0692 0.0140 0.0262 0.0000 0.0996 0.1202 0.1955 0.5637 0.0095 0.0102 0.0179
332 0.0376 0.0004 0.0017 0.0570 0.0098 0.0204 0.0000 0.0132 0.0300 0.0016 0.1912 0.0212 0.0290
331 0.2337 0.0048 0.0077 0.1556 0.0193 0.2857 0.2365 0.1983 0.0839 0.0148 0.1493 0.7535 0.0638
329 0.0325 0.0003 0.0000 0.0733 0.0148 0.0204 0.0000 0.0252 0.0191 0.0043 0.0331 0.0101 0.1385
Table 6
Citation frequency matrix and its transformation into priority matrix
455
3.01 403 7 39 73 91 130
Citation frequency matrix
329 300 0 0 2 3 30 13 152
304 0 1 3 16 38 37 347
311 0 0 0 4 3 6 27
315 0 7 0 17 64 27 471
345 0 1 0 3 2 5 30
347 0 0 0 12 24 20 129
372 0 0 0 0 0 0 1
Priority matrix
329 300 0.0000 0.0000 0.4000 0.0545 0.1863 0.1204 0.1314
304 0.0000 0.1111 0.6000 0.2909 0.2360 0.3426 0.2999
311 0.0000 0.0000 0.0000 0.0727 0.0186 0.0556 0.0233
315 0.0000 0.7778 0.0000 0.3091 0.3975 0.2500 0.4071
345 0.0000 0.1111 0.0000 0.0545 0.0124 0.0463 0.0259
347 0.0000 0.0000 0.0000 0.2182 0.1491 0.1852 0.1115
372 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0009
902 H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908
0.0000
0.0938
0.0000
0.0625
0.1875
0.0625
0.5938
0.1250
0.2500
0.0000
0.3750
0.1250
0.1250
0.0000
372
tor method. This priority vector is naturally the same as
.
.
.
.
.
.
.
.
.
.
.
the vector of the number of citations that each of 13
0.0000
0.0083
0.0062
0.0922
0.0598
0.1774
0.6561
0.1010
0.1838
0.0505
0.2929
0.0222
0.3394
0.0014
347
classes received divided by total number of citations made
.
.
.
.
.
.
.
.
.
.
.
by the patents of class 455. This is because the pairwise
0.0000
0.0208
0.0000
0.0625
0.0417
0.0208
0.8542
0.0735
0.6029
0.0441
0.1618
0.0588
0.0441
0.0020
345
comparison matrix is a completely consistent matrix.
.
.
.
.
.
.
.
.
.
.
.
Therefore, the priority vectors can be directly obtained
0.0000
0.0096
0.0032
0.0777
0.0714
0.0469
0.7913
0.0854
0.1357
0.0244
0.6113
0.0442
0.0976
0.0002
315
from the citation frequency matrix without pairwise
.
.
.
.
.
.
.
.
.
.
.
comparisons.
0.0000
0.0517
0.0000
0.1379
0.0000
0.1034
0.7069
0.0000
0.2500
0.5167
0.1000
0.0000
0.1333
0.0000
Table 5 shows the priority vectors for each cluster,
311
.
.
.
.
.
.
.
.
.
.
.
that is, the cluster weights derived in this way. The clus-
0.0011
0.0504
0.0258
0.1153
0.0582
0.0582
0.6909
0.0498
0.7658
0.0102
0.0905
0.0158
0.0645
0.0005
ter weights will be used to obtain the weighted super-
304
.
.
.
.
.
.
.
.
.
.
.
matrix.
0.0000
0.0126
0.0126
0.0755
0.0943
0.0377
0.7673
0.5038
0.1515
0.0114
0.2879
0.0303
0.0114
0.0005
Secondly, local priority vectors for mainline subclasses
329
300
.
.
.
.
.
.
.
.
.
.
.
are obtained. In ANP, basically, pairwise comparisons
are made among elements of a cluster an arrow enters with
331
respect to each element of a cluster from which an arrow
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
leaves. For a feedback loop, elements in a cluster are
332
pair-wisely compared with respect to each element in the
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
cluster itself. For each pairwise comparison supposed to
340
be made, local priority vectors can be directly derived
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
without pairwise comparisons as mentioned above. For 341
example, the importance of mainline subclasses of class .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
370 on each mainline subclass of class 450 is obtained by
transformation of the citation frequency matrix, as shown
342
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
in Table 6. What is important here is normalization of col-
umns has to be done for each cluster. The resulting set of
343
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
priority vectors, a priority matrix, will be imported to the
supermatrix.
367
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
370
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
block corresponds to a set of priority vectors, a priority
matrix. The priority matrix in Table 6 is equivalent to
380
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
W13,1 in the supermatrix. Table 7 shows a part of the limit
0.0011
0.0089
0.0086
0.0479
0.1020
0.1006
0.7310
0.1314
0.2999
0.0233
0.4071
0.0259
0.1115
supermatrix. 0.0009
130
.
.
.
.
.
.
.
.
.
.
.
0.1204
0.3426
0.0556
0.2500
0.0463
0.1852
0.0000
.
.
.
.
.
.
.
.
.
.
.
0.1863
0.2360
0.0186
0.3975
0.0124
0.1491
0.0000
0.0545
0.2909
0.0727
0.3091
0.0545
0.2182
0.0000
0.4000
0.6000
0.0000
0.0000
0.0000
0.0000
0.0000
.
.
.
.
.
.
.
.
.
.
.
0.0000
0.1111
0.0000
0.7778
0.1111
0.0000
0.0000
403
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
3.01
455
130
300
304
311
315
345
347
372
Supermatrix
.
.
.
.
.
.
.
.
.
.
.
455
380
379
375
370
367
343
342
341
340
332
331
329
455 380 379 375 370 367 343 342 341 340 332 331 329
3.01 403 7 39 73 91 130 . . . . . . . . . . . 300 304 311 315 345 347 372
455 3.01 0.0000 0.0002 0.0000 0.0017 0.0002 0.0008 0.0004 . . . . . . . . . . . 0.0000 0.0002 0.0000 0.0000 0.0000 0.0000 0.0000
403 0.0695 0.2339 0.1144 0.0312 0.0547 0.0108 0.0030 . . . . . . . . . . . 0.0026 0.0103 0.0105 0.0020 0.0043 0.0017 0.0221
7 0.0000 0.0125 0.0635 0.0078 0.0096 0.0036 0.0029 . . . . . . . . . . . 0.0026 0.0053 0.0000 0.0007 0.0000 0.0013 0.0000
39 0.1042 0.0249 0.0394 0.0953 0.0189 0.0285 0.0161 . . . . . . . . . . . 0.0154 0.0235 0.0281 0.0159 0.0128 0.0188 0.0147
73 0.1911 0.0524 0.0839 0.0295 0.0730 0.0513 0.0343 . . . . . . . . . . . 0.0193 0.0119 0.0000 0.0146 0.0085 0.0122 0.0442
91 0.0695 0.0025 0.0254 0.0321 0.0391 0.1781 0.0339 . . . . . . . . . . . 0.0077 0.0119 0.0211 0.0096 0.0043 0.0362 0.0147
130 0.0695 0.0107 0.0127 0.1395 0.1413 0.0637 0.2461 . . . . . . . . . . . 0.1570 0.1409 0.1441 0.1615 0.1748 0.1338 0.1400
380 . . . . . . . . . . . . . . . . . . . . . . . . . .
379 . . . . . . . . . . . . . . . . . . . . . . . . . .
375 . . . . . . . . . . . . . . . . . . . . . . . . . .
370 . . . . . . . . . . . . . . . . . . . . . . . . . .
367 . . . . . . . . . . . . . . . . . . . . . . . . . .
343 . . . . . . . . . . . . . . . . . . . . . . . . . .
342 . . . . . . . . . . . . . . . . . . . . . . . . . .
Table 9
Limit supermatrix
455 380 379 375 370 367 343 342 341 340 332 331 329
3.01 403 7 39 73 91 130 . . . . . . . . . . . 300 304 311 315 345 347 372
455 3.01 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 . . . . . . . . . . . 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003
403 0.0194 0.0193 0.0194 0.0194 0.0194 0.0194 0.0194 . . . . . . . . . . . 0.0193 0.0193 0.0193 0.0193 0.0193 0.0193 0.0193
7 0.0032 0.0032 0.0032 0.0032 0.0032 0.0032 0.0032 . . . . . . . . . . . 0.0032 0.0032 0.0032 0.0032 0.0032 0.0032 0.0032
39 0.0164 0.0164 0.0164 0.0164 0.0164 0.0164 0.0164 . . . . . . . . . . . 0.0164 0.0164 0.0164 0.0164 0.0164 0.0164 0.0164
73 0.0174 0.0173 0.0174 0.0174 0.0174 0.0174 0.0174 . . . . . . . . . . . 0.0174 0.0173 0.0174 0.0174 0.0174 0.0174 0.0174
91 0.0180 0.0180 0.0180 0.0180 0.0180 0.0180 0.0180 . . . . . . . . . . . 0.0180 0.0180 0.0180 0.0180 0.0180 0.0180 0.0180
130 0.0699 0.0699 0.0699 0.0699 0.0699 0.0699 0.0699 . . . . . . . . . . . 0.0699 0.0699 0.0699 0.0699 0.0699 0.0699 0.0699
380 . . . . . . . . . . . . . . . . . . . . . . . . . .
379 . . . . . . . . . . . . . . . . . . . . . . . . . .
375 . . . . . . . . . . . . . . . . . . . . . . . . . .
370 . . . . . . . . . . . . . . . . . . . . . . . . . .
367 . . . . . . . . . . . . . . . . . . . . . . . . . .
343 . . . . . . . . . . . . . . . . . . . . . . . . . .
342 . . . . . . . . . . . . . . . . . . . . . . . . . .
341 . . . . . . . . . . . . . . . . . . . . . . . . . .
340 . . . . . . . . . . . . . . . . . . . . . . . . . .
332 . . . . . . . . . . . . . . . . . . . . . . . . . .
331 . . . . . . . . . . . . . . . . . . . . . . . . . .
329 300 0.0025 0.0025 0.0025 0.0025 0.0025 0.0025 0.0025 . . . . . . . . . . . 0.0025 0.0025 0.0025 0.0025 0.0025 0.0025 0.0025
304 0.0102 0.0102 0.0102 0.0102 0.0102 0.0102 0.0102 . . . . . . . . . . . 0.0102 0.0102 0.0102 0.0102 0.0102 0.0102 0.0102
311 0.0015 0.0015 0.0015 0.0015 0.0015 0.0015 0.0015 . . . . . . . . . . . 0.0015 0.0015 0.0015 0.0015 0.0015 0.0015 0.0015
315 0.0075 0.0075 0.0075 0.0075 0.0075 0.0075 0.0075 . . . . . . . . . . . 0.0075 0.0075 0.0075 0.0075 0.0075 0.0075 0.0075
345 0.0008 0.0008 0.0008 0.0008 0.0008 0.0008 0.0008 . . . . . . . . . . . 0.0008 0.0008 0.0008 0.0008 0.0008 0.0008 0.0008
347 0.0026 0.0026 0.0026 0.0026 0.0026 0.0026 0.0026 . . . . . . . . . . . 0.0026 0.0026 0.0026 0.0026 0.0026 0.0026 0.0026
372 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 . . . . . . . . . . . 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
903
904 H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908
Table 10 from the USPTO, the ANP network model was con-
Limit centralities of 13 classes structed and local priority vectors were obtained. Forming
No. Title Limit and transforming the supermatrix led to converged priori-
centrality ties, limit centralities.
455 Telecommunications 0.0251 The main contribution of this study is to apply the
380 Cryptography 0.2885 MCDM methodology, ANP, to a technology network.
379 Telephonic communications 0.0272
375 Pulse or digital communications 0.0883
Since ANP captures the relative importance that mirrors
370 Multiplex communications 0.0191 all the direct and indirect interactions, the limit centrality
367 Communications, electrical: acoustic wave systems 0.0219 measures importance of technologies in terms of impacts
and devices on other technologies in the technology network, taking
343 Communications: radio wave antennas 0.0034 indirect impacts or relationships into account, which is very
342 Communications: directive radio wave systems and 0.0023
devices
difficult or tedious with the conventional centrality mea-
341 Coded data generation or conversion 0.0515 sures. The applicability of limit centrality is not limited
340 Communications: electrical 0.1547 to a technology network. For any type of social networks,
332 Modulators 0.0263 the limit centrality can be used as an implicative centrality
331 Oscillators 0.1464 measure characterizing a network and showing core actors
329 Demodulators 0.1446
in the network.
Nevertheless, this research is still subject to some limi-
tations. Firstly, it cannot be used for undirectional net-
4.5. Limit centrality works where an edge has no direction and only
represents the existence of a relationship between two
As the supermatrix covers the whole network, the col- nodes since relationships in a network of ANP must have
umns in the limit supermatrix (Table 9) represent final pri- directions depending on the influence between elements or
orities, namely, limit centralities. That is why it is called clusters. Secondly, the influences among patent classes are
limit centrality. Due to the nature of limit priorities in measured by the absolute size of patent citations; thus, we
ANP, the limit centralities of all the elements sum to one. cannot control the effect of the size of a class, that is, total
The limit centrality indicates importance of technologies number of patents in a class, on measuring the degree of
in terms of impacts on other technologies, taking all the impacts. The relative impact may be a more implicative
direct and indirect influences into consideration. The limit measure depending on the context. It can be derived by
centralities of the 159 mainline subclasses are shown in dividing each column of the citation frequency matrix
Appendix B. The limit centrality of a class is the sum of by the total number of patents in the corresponding class.
mainline subclasses belonging to the class. Table 10 shows Thirdly, the selected 13 patent classes as telecommunica-
the limit centralities of 13 classes. tion technologies are by no means exhaustive. A more
At the mainline subclass level, the one with the highest systematic procedure is required to select the target
limit centrality is 455/130 (Receiver or analog modulated classes.
signal frequency converter), and the next is 375/316 These limitations could serve as fruitful avenues for
(Receivers). It is obvious that these technologies have sig- future research. Applications of the proposed approach
nificant impacts on other technologies, and therefore they to a variety of networks can be a worthwhile area for future
are considered as the core technologies of the telecommuni- research. A dynamic analysis on the telecommunication
cation technology network. When it comes to the class network is also expected to provide useful information on
level, 380 (Cryptography) has the highest limit centrality, the change of the network structure and technological
followed by 340 (Communications: electrical), 331 (Oscilla- trends.
tors), and 329 (Demodulators).
On the other hand, the limit centrality of 379/414 (trans-
mission line conditioning) is zero since its patents have
never been cited by all the patents of the other classes. Appendix A. ICT classification and corresponding IPC codes
The class whose limit centrality is the lowest is 342 (Com-
munications: directive radio wave systems and devices).
Classifications IPC codes
5. Conclusions Telecommunication G01S, G08C, G09C, H01P, H01Q,
H01S3/(025, 243, 063, 067, 085,
The proposed approach measures the limit centralities 0933, 0941, 103, 133, 18, 19, 25),
of technologies with the aim of identification of core tech- H1S5, H03B, H03C, H03D, H03H,
nologies in the technology network. A case study on the H03M, H04B, H04J, H04K, H04L,
telecommunication technology network was presented to H04M, H04Q
illustrate the proposed approach. After constructing the Consumer G11B, H03F, H03G, H03J, H04H,
citation frequency matrix based on patent data collected electronics H04N, H04R, H04S
H. Lee et al. / Expert Systems with Applications 36 (2009) 894–908 905
Ernst, H. (2003). Patent information for strategic technology manage- Lin, Y. H., Chiu, C. C., & Tsai, C. H. (2007). The study of applying ANP
ment. World Patent Information, 25(3), 233–242. model to assess dispatching rules for wafer fabrication. Expert Systems
Freeman, L. C. (1979). Centrality in social networks: Conceptual with Applications, doi:10.1016/j.eswa.2007.02.033.
clarification. Social Networks, 1, 215–239. Marseden, P., & Laumann, E. (1984). Mathematical ideas in social
Gangulli, P. (2004). Patents and patent information in 1979 and 2004: A structural analysis. Journal of Mathematical Sociology, 10, 271–294.
perspective from India. World Patent Information, 26, 61–62. Meade, L., & Presley, A. (2002). R&D project selection using ANP. IEEE
Gelsing, L. (1992). Innovation and the development of industrial Transactions on Engineering Management, 49(1), 22–28.
networks. In B.-A. Lundvall (Ed.), National systems of innovation – Meade, L., & Sarkis, J. (1998). Strategic analysis of logistics and supply
Towards a theory of innovation and interactive learning (pp. 116–128). chain management systems using the analytical network process.
London: Pinter. Transportation Research Part E: Logistics and Transportation Review,
Gencer, C., & Gürpinar, D. (2007). Analytic network process in supplier 34(3), 201–215.
selection: A case study in an electronic firm. Applied Mathematical Meade, L., & Sarkis, J. (1999). Analyzing organizational project alterna-
Modelling, 31(11), 2475–2486. tives for agile manufacturing processes: An analytic network approach.
Grilliches, Z. (1990). Patent statistics as economic indicators: A survey. International Journal of Production Research, 37(2), 241–261.
Journal of Economic Literature, 28, 1661–1707. Mowery, D. C., Oxley, J. E., & Silverman, B. S. (1998). Technological
Grupp, H. (1996). Spillover effects and the science base of innovations overlap and interfirm cooperation: Implications for the resource-based
reconsidered: An empirical approach. Journal of Evolutionary Eco- view of the firm. Research Policy, 27, 507–523.
nomics, 6(2), 175–197. Mulebeke, J. A. W., & Zheng, L. (2006). Analytical network process for
Grupp, H., Lacasa, D., & Schmoch, U. (2003). Tracing technological software selection in product development: A case study. Journal of
change over long periods in Germany in chemicals using patent Engineering and Technology Management, 23(4), 337–352.
statistics. Scientometrics, 57, 175–195. Narin, F. (1994). Patent bibliometrics. Scientometrics, 30, 147–155.
Gun̈gor̈, A. (2006). Evaluation of connection types in design for Narin, F., Noma, E., & Perry, R. (1987). Patents as indicators of
disassembly (DFD) using analytic network process. Computers & corporate technological strength. Research Policy, 16, 143–155.
Industrial Engineering, 50(1/2), 35–54. Niemira, M. P., & Saaty, T. L. (2004). An Analytic Network Process
Hall, B. H., Jaffe, A. B., & Trajtenberg, M. (2001). The NBER patent model for financial-crisis forecasting. International Journal of Fore-
citations data file: Lessons, insights and methodological tools. casting, 20(4), 573–587.
National Bureau of Economic Research, Working Paper No. 8498. OECD (2005). Measuring globalization: OECD handbook on economic
Ham, R. M., Linden, G., & Appleyard, M. M. (1998). The evolving role of globalization indicators. OECD, Paris.
semiconductor consortia in the United States and Japan. California Reitzig, M. (2004). Improving patent valuations for management pur-
Management Review, 41, 137–163. poses—validating new indicators by analyzing application rationales.
Han, Y., & Park, Y. (2006). Patent network analysis of inter-industrial Research Policy, 33(6/7), 939–957.
knowledge flows: The case of Korea between traditional and emerging Ruhnau, B. (2000). Eigenvector-centrality — a node-centrality? Social
industries. World Patent Information, 28(3), 235–247. Networks, 22(4), 357–365.
Harhoff, D., Scherer, F. M., & Vopel, K. (2003). Citations, family size, Saaty, T. (1980). The analytic hierarchy process. New York: McGraw-Hill.
opposition and the value of patent rights. Research Policy, 32(8), Saaty, T. (1996). Decision making with dependence and feedback: The
1343–1363. analytic network process. Pittsburgh: RWS Publications.
Hirschey, M., & Richardson, V. J. (2001). Valuation effects of patent Schapper, M. (2003). A proposal for a core list of indicators for ICT
quality: A comparison for Japanese and US firms. Pacific-Basin measurement. Electronic document at http://www.oecd.org, OECD.
Finance Journal, 9(1), 65–82. Shin, J., & Park, Y. (2007). Building the national ICT frontier: The case of
Jaffe, A., & Trajtenberg, M. (1998). International knowledge flows: Korea. Information Economics and Policy, 19(2), 249–277.
Evidence from patent citations. National Bureau of Economic Trajtenberg, M. (1990). A penny for your quotes: Patent citations and the
Research, Working Paper No. 6509. value of inventions. Rand Journal of Economics, 21, 172–187.
Jharkharia, S., & Shankar, R. (2007). Selection of logistics service Trajtenberg, M., Henderson, R., & Jaffe, A. B. (1997). University versus
provider: An analytic network process (ANP) approach. Omega, 35(3), corporate patents: A window on the basicness of invention. Economics
274–289. of Innovation and new technology, 5(1), 19–50.
Kahraman, C., Ertay, T., & Büyüközkan, G. (2006). A fuzzy optimization USPTO (2006). Overview of the US patent classification system (USPC).
model for QFD planning process using analytic network approach. Electronic document at http://www.uspto.gov, USPTO.
European Journal of Operational Research, 171(2), 390–411. Wartburg, I., Teichert, T., & Rost, K. (2005). Inventive progress measured
Karsak, E. E., Sozer, S., & Alptekin, S. E. (2003). Product planning in by multi-stage patent citation analysis. Research Policy, 34, 1591–1607.
quality function deployment using a combined analytic network Wasserman, S., & Faust, K. (1994). Social network analysis. Cambridge
process and goal programming approach. Computers & Industrial University Press.
Engineering, 44(1), 171–190. Wei, W. L., & Chang, W. C. (2007). Analytic network process-based
Kim, Y. G., Suh, J. H., & Park, S. C. (2007). Visualization of patent model for selecting an optimal product design solution with zero-one
analysis for emerging technology. Expert Systems with Applications, goal programming. Journal of Engineering Design, doi:10.1080/
doi:10.1016/j.eswa.2007.01.033. 09544820601186054.
Lai, K. K., & Wu, S. J. (2005). Using the patent co-citation approach to Wu, W. W., & Lee, Y. T. (2007). Selecting knowledge management
establish a new patent classification system. Information Processing & strategies by using the analytic network process. Expert Systems with
Management, 41(2), 313–330. Applications, 32(3), 841–847.
Lanjouw, J. O., & Schankerman, M. (1999). The quality of ideas: Yoon, B., & Park, Y. (2004). A text-mining based patent network:
Measuring innovation with multiple indicators. National Bureau of Analytical tool for high-technology trend. Journal of High Technology,
Economic Research, Working Paper No. 7345. 15(1), 37–50.
Lee, J. W., & Kim, S. H. (2000). Using analytic network process and goal _ & Dag66deviren, M. (2007). Using the analytic network process
Yüksel, I.,
programming for interdependent information system project selection. (ANP) in a SWOT analysis – A case study for a textile firm.
Computers & Operations Research, 27(4), 367–382. Information Sciences, 177(16), 3364–3382.