You are on page 1of 10

Pattern Recognition Letters 19 1998.

237246

A new cluster validity index for the fuzzy c-mean


)
M. Ramze Rezaee, B.P.F. Lelieveldt, J.H.C. Reiber
Diision of Image Processing, Department of Radiology, Leiden Uniersity Medical Center, P.O. Box 9600, 2300 RC Leiden, The
Netherlands
Received 17 June 1997; revised 16 October 1997

Abstract

In this paper a new cluster validity index is introduced, which assesses the average compactness and separation of fuzzy
partitions generated by the fuzzy c-means algorithm. To compare the performance of this new index with a number of
known validation indices, the fuzzy partitioning of two data sets was carried out. Our validation performed favorably in all
studies, even in those where other validity indices failed to indicate the true number of clusters within each data set. q 1998
Elsevier Science B.V. All rights reserved.

Keywords: Cluster validity functional; Fuzzy clustering; Fuzzy c-mean

1. Introduction Since clustering algorithms are unsupervised, irre-


spective of the clustering method hard or fuzzy., the
The objective of most clustering methods is to final partitions of data require some kind of valida-
provide useful information by grouping unlabelled. tion in most applications. Backer and Jain 1981.
data in clusters; within each cluster the data exhibits have studied the performance of several clustering
similarity. Similarity is defined by a distance mea- techniques based on a fuzzy set decomposition of the
sure, and global objective functional or regional data set under consideration. In general, cluster vali-
graph-theoretic criteria are optimized to find the dation may answer, among other aspects, questions
optimal partitions of data. The partitions generated such as: How good are the partitions? Is there a
by a clustering approach define for all data elements better partitioning possible?, etc. In addition, if the
to which class cluster. they belong. The partitions number of classes within the data is not known a
may define a hard boundary between sub-partitions; priori, a validation index may help to find out the
this is called hard clustering. In contrast, the bound- optimal number of classes. This is performed in three
aries between sub-partitions generated by a fuzzy stages. Firstly, all parameters of the clustering
clustering algorithm are vague. This means that each method, except for the number of clusters, are fixed.
pattern of object data of a fuzzy partition belongs to Secondly, by varying the number of clusters between
different classes with different membership values. 2 and an upper value c max and applying a clustering
algorithm, for each number of clusters c i g
 2,3, . . . , c max 4 a different partition of the data is
)
Corresponding author. E-mail: hreiber@lkeb.azl.nl. found. In the final stage a validation index is applied

0167-8655r98r$19.00 q 1998 Elsevier Science B.V. All rights reserved.


PII S 0 1 6 7 - 8 6 5 5 9 7 . 0 0 1 6 8 - 2
238 M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246

to each partition, obtained at the second stage, to represents the grade of membership of data point x k
define a validation value. The true number of data of set X s  x 1 , x 2 , . . . , x n 4 to the ith cluster. The
classes can be determined based on extrema of these inner product defined by a norm matrix A defines a
validation values for all c i . measure of similarity between a data point and the
In this paper we will introduce a new validation cluster prototypes. A nondegenerate fuzzy c-partition
index, which measures the separation between clus- of X is conveniently represented by a matrix U s
ters and the cohesion within clusters, which are w u i k x.
obtained by the fuzzy c-mean FCM.. After a brief It has been shown by Bezdek 1981. that if
review of the FCM in Section 2, a number of 5 x k y i 5 A ) 0 for all i and k, then U,V . may
validation indices is described in Section 3. In addi- minimize Jm only, when m ) 1 and
tion, in this section the new validation index is n
m
introduced. To evaluate the performance of this vali- uik . x k
dation index, one synthetic and one real world data ks1
i s n for 1 ( i ( c, 2.
set are used and the optimal number of object classes m
is assessed by applying a number of validation in- uik .
ks1
dices, including the new index. The results of the
evaluation study are described in Section 5. Finally, 1
uik s
this paper concludes with a discussion section Sec- 1rmy1
c 5 x k y i 5 2A
tion 4..
js1 5 x k y j 5 2A /
2. Fuzzy c-mean algorithm for 1 ( i ( c, 1 ( k ( n. 3.
Among others, Jm can be minimized by the Pi-
Fuzzy c-mean FCM. is an unsupervised cluster-
card iteration approach. This method minimizes Jm
ing algorithm that has been applied successfully to a
by initializing the matrix U randomly or predefined.
number of problems involving feature analysis, clus-
and computing the cluster prototypes Eq. 2.. and
tering and classifier design. FCM has a wide domain
the membership values Eq. 3.. after each iteration.
of applications such as agricultural engineering, as-
The iteration is terminated when it reaches a stable
tronomy, chemistry, geology, image analysis, medi-
condition. This can be defined for example, when the
cal diagnosis, shape analysis and target recognition
Bezdek, 1987.. Unlabeled data are classified by changes in the cluster centers or the membership
values at two successive iteration steps is smaller
minimizing an objective function based on a norm
than a predefined threshold value. The FCM algo-
and clusters prototype. Although the description of
rithm always converges to a local minimum or a
the original algorithm dates back to 1973 Bezdek,
saddle point. A different initial guess of u i j may
1973; Dunn, 1974. derivatives have been described
lead to a different local minimum. Finally, to assign
with modified definitions for the norm and proto-
each data point to a specific cluster, defuzzification
types for the cluster centers Dave and Bhaswan,
is necessary, e.g., by attaching a data point to a
1992; Krishnapuram et al, 1992; Man and Gath,
cluster for which the value of the membership is
1994..
maximal.
The FCM minimizes an objective function Jm ,
which is the weighted sum of squared errors within
groups and is defined as follows:
n c 3. Validation indices for the fuzzy c-mean
Jm U,V ; X . s u imk 5 x k y i 5 2A , 1 - m - `,
ks1 is1
3.1. Validation criteria
1.
where V s 1 , 2 , . . . , c . is a vector of unknown Although the FCM can find a partition of data for
cluster prototype centers. i g R p. The value of u i k a fixed number of clusters objects., one objective of
M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246 239

a cluster validity procedure is to determine automati- Pal and Bezdek, 1995.; the best partition is achieved
cally the optimal number of clusters. This is desired, when the value for V PC has a maximum or the value
for example, when the FCM is used for image for V PE has a minimum, for a certain number of
segmentation purposes and the number of objects in clusters. However, a strong criticism against the
the image is not known a priori. Validation of a indices V PC or VPE , is that they are only implicitly a
generated fuzzy partition by the FCM can be achieved function of the data set X. It means that they do not
by a cluster validity index. If c min is defined as the use the data itself. To overcome this lack of direct
minimum and c max as the maximum number of connection to the geometrical properties of the data
clusters, respectively, then for each c g w c min , c max x a set, Gundersons separation coefficient Gunderson,
partition can be generated by the FCM. The value of 1978. uses both the data set and the prototypes
a cluster validity index can then be calculated for the cluster centers.. This is also the case for the
partitions of data for each c. By comparing all values Fukuyamas and Sugenos index Fukuyama and
of an index for all possible number of clusters, one Sugeno, 1989. and the XieBenis index Xie and
can determine the optimal number of clusters. This Beni, 1991.. A limited analysis of these indices was
can be achieved, for example, by selecting the num- performed by Pal and Bezdek 1995, 1997.. They
ber of clusters for which an index is minimized. also analyzed the reliability of these indices as a
There are a number of clusters validation indices function of the weighting exponent m of Eq. 1..
available Bezdek, 1974, 1975; Gunderson, 1978;
Windham, 1981, 1982; Libert and Roubens, 1983; 3.2. Proposed alidation index
Windham et al., 1989; Fukuyama and Sugeno, 1989;
Xie and Beni, 1991; Gindy et al., 1995.. Some A reliable validation functional for the FCM must
validity methods use only the membership values of consider both the compactness and the separation of
a fuzzy partition of data. Among other functional, a fuzzy c-partition. The optimal partitions require a
such indices are the partition coefficient V PC Be- maximal compactness for each cluster partition. in
zdek, 1974., the partition entropy VPE Bezdek, such a way that the clusters are located far from each
1975., the proportion exponent Windham, 1981. other. If only the compactness requirement is consid-
and the uniform data functional Windham, 1982.. ered by a validation functional, the best partition is
Table 1 lists a number of cluster validation in- obtained when each data point is considered as a
dices, which are evaluated in our study. The func- separate cluster; nothing is compacter than a cluster
tional partition coefficient V PC and the partition en- that includes only one data point. On the other hand,
tropy VPE use only the membership values u i k of a if only the optimal separation between the clusters is
fuzzy partition of data set X. Some empirical studies considered as the single validation criterion, the best
have shown that maximizing VPC or minimizing partition will be the data set itself; the distance
VPE . often leads to a good interpretation of the data between a cluster the total data set. and itself is

Table 1
Four validation functionals for the fuzzy c-mean
Validity index Functional description Optimal cluster number
1
Partition coefficient VPC U . s nks1 cis1 u 2i k . MaxVPC U,c i ,m.4
n
1
Partition entropy VPE U . s y nks1 cis1 u i k log a u i k .. MinV PE U,c i ,m.4
n 2
Fukuyama and Sugeno 1989. VFS , m U,V ; X . s cis1 nks1 u imk 5 x k y i 5 2 y i y A
. MinV FS U,c i ,m.4
cis 1 nks1 u imk 5 x k y i 5 2
Xie and Beni 1991. VXB U;V; X . s MinV XB U,c i ,m.4
n min  i y j 4 .

x k is the k th data point, i are cluster prototypes cluster centers., c i is the number of clusters, is the grand mean of all data x k and u i k is
the membership value of data x k of class c i .
240 M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246

zero. Therefore a reliable validation functional will as Dmax , but for the minimum distance between the
have an optimal value for that partition, which com- cluster prototypes.
bines both criteria. In an attempt to find such a
functional, we have designed a new validation index Our validation index VCW B is now defined by
which includes the compactness and separation cri- combining the last two equations:
teria. We called this index VCW B Compose Within
VCW B U,V . s a Scat c . q Dis c . , 8.
and Between scattering.. To define this functional,
let us first describe the following requirements and where a is a weighting factor equal to Dis c max ..
definitions. The first term of VCW B , i.e. Scat c . of Eq. 6.,
indicates the average of the scattering variation.
Requirements. A fuzzy c-partition of the data set within the clusters for c number of clusters. A small
X s  x 1 , x 2 , . . . , x n < x i g R p 4 with c cluster centers value for this term indicates a compact partition. As
i , such that V s  1 , 2 , . . . , c 4 and U s w u i k i s the scattering within the clusters increases, they be-
1,2, . . . , c; k s 1,2, . . . , n.x. come less compact, and therefore Scat c . is a good
indication of the average compactness of clusters. In
Definition 1. The variance of the pattern set X is general, the scattering index Scat c . does not take
called s X . g R p with the value of the pth dimen- any geometric assumption about the prototype into
sion defined as account. The second term of our validation index,
n
Dis c ., indicates the total scattering separation. be-
1 2 tween the clusters. Generally, this term will increase
sx p s x kp y x p . , 4.
n ky1 with the number of clusters and is influenced by the
geometry of the cluster centers. Since the values of
where x p is the pth value of the grand mean of
the two terms of VCW B are of a different range, a
X s nks 1 x krn, ; x k g X.
weighting factor a is needed in order to counterbal-
ance both terms in a proper way.
Definition 2. The fuzzy variation of the cluster i is A cluster number, which minimizes the validation
called s i . g R p with the pth value defined as index VCW B can be considered as an optimal value
1 n
2 for the number of object classes present in the data.
s ip s u i k x kp y ip . . 5.
n ks1
3.3. FCM alidation algorithm

Definition 3. The average scattering for c clusters is If the minimum and maximum values of the
defined as number of clusters be denoted as c min and cmax ,
1 c respectively, then by the following pseudo algorithm
c
5 s i . 5 an optimal number of clusters c opt of a data set can
is1 be obtained.
Scat c . s , 6.
5s X . 5 1. Initialization: c opt s c c max ;
where 5 x 5 s x T P x .1r2 . 2. Apply FCM to the data set to update the cluster
centers i and the membership values m i k
3. Do iteration and test for convergence; if not
Definition 4. A distance functional Dis c . is defined
goto 2.
as
4. if c s c max .  a s Dis c max .; indexValue s
y1
Dmax c c Vcwb c .4
Dis c . s
Dmin

ks1
zs1
5 k y z 5
/ , 7. else if VCW B c . - indexValue.  c opt c; in-
dexValues VCW B c .;4
where Dm ax s maximum  5 i y j 54 ; i, j g 5. c c y 1, if c s c min y 1. stop else goto 2.
 2,3, . . . , c4 is the maximum distance between the In step 3 the convergence of FCM can be tested, e.g.,
cluster prototypes. The Dmin has the same definition by the condition < u i k t q 1. y u i k t .< - e ; i g
M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246 241

w c min , c max x n ;k g  1,2, . . . , n4 , where t and t q 1 and Glesner 1994. have shown that a very good
represent two successive iteration steps. After step 3 classification can be obtained by using only two
the fuzzy partition is generated. This partition is features. As demonstrated in the next section, a very
validated by the VCW B in step 4. The parameter good classification rate can be obtained when only
indexValue stores the minimum value of VCW B the feature petal length is used.
so far. After step 5 the optimal number of clusters
c opt will be found that corresponds to the minimum Parameters of the study. In our study the parame-
value of the index VCW B within the range of ters of the FCM were set as follows: the stop crite-
w c min , c max x. rion for the iteration was e s 0.001. The norm 5 ) 5 A
Also, by varying the fuzzy parameter m g 1,`. was the Euclidean and all iterations began with a
of the objective function of Eq. 1. and by applying randomly initialized u i k . Two parameters of the
the above algorithm, the optimal number of clusters FCM, the number of clusters c and the weighting
is found that is the function of VCW B m,c .. This exponent m, were varied for both data sets. The
optimum minimizes the cluster validity VCW B . weighting exponent m was varied from 1.1 to 2.0
with a step value of 0.05 total of 19 samples.. To
obtain the optimal number of clusters for each fixed
4. Data acquisition and FCM parameters value of m, nine different partitions were made by
varying the number of clusters from two to ten and
To compare the performance of the proposed applying the FCM. All these nine partitions were
validation index with a number of known validation then used to calculate the value of five validation
indices, an evaluation study was carried out. This indices: the partition coefficient VPC , the partition
study included two data sets. entropy V PE , Xie and Benis index VXB , Fukuyamas
and Sugenos index V FS and the VCWB . By compar-
Normal-4. The first data set called Normal-4 was ing the values of the validation index of a specific
described by Pal and Bezdek 1995. as follows: functional V m s fixed, 2 ( c ( 9., the optimal
Normal-4 includes 800 data points consisting of number of clusters was defined by that functional. In
clusters of 200 points, each of the four components total 171 partitions 19 samples for m = 9 possibili-
of a mixture of c s 4, p s 4-variate normal. The ties for the number of clusters. were generated for
population mean vector and covariance matrix for each data set by applying the FCM. The optimal
each component of the normal mixture were m i s 3e i number of clusters for each specific m value was
and S i s I4 , i s 1,2,3,4, and e i s 0, . . . , 1, . . . , 0. then registered.
the 1 is on the ith position..

IRIS. The second data set of the evaluation study 5. Results


was the Iris data set of AndersonFisher Anderson,
1935; Fisher, 1936.. This is a biometric data set Normal-4. For 4 number of clusters the average
consisting of 150 measurements belonging to three classification error of FCM was 3.5% for all 19
flower varieties: Setosa, Versicolor and Virginica, samples of m. Fig. 1ae. shows the validation
generally known as the Iris data set. Each class index values of all indices as a function of the
includes 50 observations, in which two variables, weighting exponent m and the number of clusters c.
length and width of the petal and sepal, are mea- As Fig. 1a. demonstrates, by increasing the value of
sured. Since the length and the width of each vari- the weighting exponent m the partition coefficient
able are measured, each individual measurement one V PC m,c . decreases. Up to the value of m s 1.75,
flower out of 150 samples. is represented as a point for each fixed m the maximum of V PC m,c . indi-
in p s 4-dimensional measurement space. Pal and cates 4 as the true numbers of clusters correctly. For
Bezdek 1997. have indicated that since two of the m 0 1.8 the maximum value of V PC m,c . appears at
three classes have a substantial overlap, one can VPC m,c s 2., indicating 2 as the optimal number of
argue in favor of both c s 2 and c s 3. Halgamuge clusters. Fig. 1b. shows the validation index entropy
242 M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246

Fig. 1. The validation index of 4-Normal data set as a function of the number of clusters c and the weighting exponent m for: a. partition
coefficient V PC ; b. partition entropy VPE ; c. Fukuyama and Sugenos index VFS ; d. Xie and Benis index VXB ; e. the index VCWB . The
grey scales in the figures indicate the different validating index values.
M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246 243

Fig. 1 continued..
244 M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246

Fig. 1 continued..

Fig. 2. The validation index VCW B for the IRIS data set.
M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246 245

V PE c,m.. In contrast to VPC by increasing the value V m,c . of all five indices was found. Since the
of the weighting exponent m this index increases. weighting exponent m had 19 samples, a reliable
Up to the value of m s 1.55 the minimum of this validation index must find 19 times the true number
index for a specific sample of m and the variable of clusters. This is indeed the case for our validation
numbers of clusters c g  2,3, . . . , 104. appears at index VCW B on the last row of Table 2: for all
V PE m,c s 4., indicating 4 as the optimal number of samples of m, three is indicated as the optimal
clusters. For all samples of m with a value m 0 1.55, number of clusters. This is the minimum value of
the VPE m,c s 2. is the minimum of this index de- VCW B m,c . for each sample of m. The VXB indicated
noting 2 as the optimal number of clusters incor- 3 to be the correct number of clusters for 9 samples
rectly. The minimal values of the validation index of of 1.30 ( m and m 0 1.85. This can be found in the
the FukuyamaSugeno index V FS m,c . of Fig. 1c. second column of the third row of the same table.
indicates 4 as the optimal number of clusters for For all other ten samples this index indicated 2 to be
m 0 1.7. The minimum of this index appears, how- the correct numbers of clusters the first column of
ever, at V FS m,10. for 1.1 ( m ( 1.65, again denot- the third row.. Fukuyamas and Sugenos index V FS
ing 10 as the true number of clusters incorrectly. Fig. indicates the higher number of clusters as the opti-
1d. shows the validation index V XB m,c .. The mini- mal value for the number of clusters.
mum values of this index appear at V XB m,4. for all
19 samples of m, which indicates the optimal num-
6. Discussion
ber of clusters to be equal to 4. Also the minimum of
our validation index VCW B m,c ., presented in Fig. In this study we have used two data sets to
1e., appears at VCW B m,4. for all samples of the evaluate the performance of a number of known
weighting exponent m. validation functional. In the context of this study, a
validation index is called reliable when the optimal
IRIS data set. For c s 3 all feature combinations number of clusters indicated by a validation index
2 4 y 1 s 15. were used to assess which one has the was equal to the true number of clusters of a data set.
smallest classification error. This was the case for To find this number, nine fuzzy partitions for each
the petal width. When this feature was used for all data set were obtained, by fixing the weighting
iris samples, the average classification error of the exponent m, varying the number of clusters from 2
FCM was 6% 9 wrongly classified samples.. to 10 and applying the FCM. These partitions were
From Fig. 2 it is clear that for all samples of the then used to obtain the validation values V m,c . for
weighting exponent m, the minimum appears at each of the five validation indices. The results of the
VCW B m,3. indicating 3 to be the true number of first data set of 4-Normal classes show the effect of
clusters. the weighting exponent m on each validation index
Table 2 presents the results for the IRIS data set. Fig. 1ae... Some validation indices such as
To obtain this table for each sample of weighting V PC m,c . or VFS m,c . decrease when the weighting
exponent m the optimal value of validation index exponent m increases. In contrast, other indices in-
crease with an increasing value of m. This directly
effects the local extrema minimum or maximum. of
Table 2 three validation indices V PC , V PE and V FS for a fixed
Results of five validation indices by using the petal width of IRIS
data
m value: in case of V PC for sample values of
m 0 1.8 the maximum of V PC m,c . shifted from the
Number of clusters
point V PC m,c s 4. to V PC m,c s 2.. Therefore, for
2 3 4 5 6 7 8 9 10 m 0 1.8 the maximum of V PC did not indicate the
Partition coefficient 6 7 0 0 0 0 2 2 2 true numbers of clusters. This was also the case for
Partition entropy 9 5 0 0 0 0 2 1 2 V PE m,c ., for each sample of m 0 1.55 the mini-
Xie and Beni 1991. 10 9 0 0 0 0 0 0 0
Fukuyama and Sugeno 1989. 0 0 0 1 2 0 2 4 10
mum was shifted from VPE m,c s 4. to V PE m,c s
VCW B 0 19 0 0 0 0 0 0 0 2.. The minimum of VFS m,c . shifted from V FS m,c
s 4. to V FS m,c s 10.. On the other hand, the mini-
246 M. Ramze Rezaee et al.r Pattern Recognition Letters 19 (1998) 237246

mum of the VXB and our validation VCWB remained Bezdek, J.C., 1974. Cluster validity with fuzzy sets. J. Cybernet. 3
3., 5872.
unchanged at VFS m,c s 4. for all 19 samples of m.
Bezdek, J.C., 1975. Mathematical models for systematics and
The results of the IRIS data set also demonstrated taxonomy. In: Estabrook, G. Ed.., Proc. 8th Internat. Conf.
that VCW B was the most reliable index. According to Numerical Taxonomy. Freeman, San Francisco, CA, pp. 143
this index, for all samples of m within the range 166.
w1.1,2x, three classes exist within the IRIS data. Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective
The results of this study suggest that the new Function Algorithms. Plenum, New York.
Bezdek, J.C., 1987. Partition structures: A tutorial. In: Bezdek,
validation index can achieve the optimal result for J.C. Ed.., The Analysis of Fuzzy Information. CRC Press,
any possible data set. This is, however, not claimed Boca Raton, FL.
here. Every validation index may fail, for example, Dave, R.N., Bhaswan, K., 1992. Adaptive fuzzy c-shells cluster-
when the numerical representation chosen to de- ing and detection of ellipses. IEEE Trans. Neural Networks 3
5., 643662.
scribe the different object features do not properly
Dunn, J.C., 1974. A fuzzy relative of the ISODATA process and
discriminate between different classes. Also the Eu- its use in detecting compact, well-separated clusters. J. Cyber-
clidean norm used in VCW B . may be unreliable for a net. 3, 3257.
specific data set. In addition, the weighting exponent Fisher, R.A., 1936. The use of multiple measurements in taxo-
m and also the random initial fuzzy partition of data nomic problems. Annals of Eugenics 7 II., 179188.
may affect the reliability of a validation index. Fukuyama, Y., Sugeno, M., 1989. A new method of choosing the
number of clusters for the fuzzy c-mean method. In: Proc. 5th
Because of practical implications, we have ap- Fuzzy Syst. Symp., pp. 247250 in Japanese..
plied the FCM by taking only a few samples of the Gindy, N.N.Z., Ratchev, T.M., Case, K., 1995. Component group-
weighting exponent m up to the value m s 2. This ing for GT applications a fuzzy clustering approach with
parameter is, however, a continuous variable in the validity measure. Internat. J. Prod. Res. 33 9., 24932509.
range 1,`.. Although a careful analysis of VCW B is Gunderson, R., 1978. Application of fuzzy ISODATA algorithms
to star-tracker pointing systems. In: Proc. 7th Triannual World
required for very large values of m, many applica- IFAC Congr., Helsinki, pp. 13191323.
tion of FCM have been realized with a weighting Halgamuge, S.K., Glesner, M., 1994. Neural networks in design-
exponent m s 2, which is included in our study. ing fuzzy systems for real world applications. Fuzzy Sets and
However, we realize that a more rigorous and Systems 65 1., 112.
theoretical study of VCW B is necessary, especially to Krishnapuram, R., Nasraoui, O., Keller, J., 1992. The fuzzy c
spherical shells algorithm: A new approach. IEEE Trans.
analyze the effect of the weighting factor a in the Neural Networks 3 5., 663671.
performance of VCW B . In addition, to establish the Libert, G. Roubens, M. 1983. New experimental results in cluster
performance of this index, testing on other challeng- validity of fuzzy clustering algorithms. In: Janssen, J., Macro-
ing synthetic and real world data seems necessary. torchino, J.F., Proth, J.M. Eds.., New Trends in Data Analy-
sis and Applications. North-Holland, Amsterdam, pp. 205218.
Man, Y., Gath, I., 1994. Detection and separation of ring-shaped
Acknowledgements clusters using fuzzy clustering. IEEE Trans. Pattern Anal.
Machine Intell. 16 8., 855861.
We wish to thank the referees for their valuable Pal, N.R., Bezdek, J.C., 1995. On cluster validity for the fuzzy
c-means model. IEEE Trans. Fuzzy Syst. 3 3., 370379.
comments and suggestions. This research is sup- Pal, N.R., Bezdek, J.C., 1997. Correction to on cluster validity for
ported by the Science Foundation, The Netherlands, the fuzzy c-means model. IEEE Trans. Fuzzy Syst. 5 1.,
project LGN.22.2781. 152153.
Windham, M.P., 1981. Cluster validity for fuzzy clustering algo-
rithms. Fuzzy Sets Syst. 5, 177185.
References Windham, M.P., 1982. Cluster validity for the fuzzy c-means
clustering algorithm. IEEE Trans. Pattern Anal. Machine In-
Anderson, E., 1935. The irises of the Gaspe Peninsula. Bull. tell. 4 4., 357363.
Amer. Iris Soc. 59, 25. Windham, M.P., Bock, H., Walker, H.F., 1989. Clustering infor-
Backer, E., Jain, A.K., 1981. A clustering performance measure mation from convergence rate. In: Proc. 2nd Conf. Internat.
based on fuzzy set decomposition. IEEE Trans. Pattern Anal. Federation Classification Soc., Washington, DC, p. 143.
Machine Intell. 3 1., 6675. Xie, X.L., Beni, G.A., 1991. A validity measure for fuzzy cluster-
Bezdek, J.C., 1973. Fuzzy mathematics in pattern classification. ing. IEEE Trans. Pattern Anal. Machine Intell. 13 8., 841
Ph.D. dissertation, Cornell University, Ithaca, NY. 847.

You might also like