Professional Paper

Comparison of Fuzzy c-means Algorithm and New

Fuzzy Clustering and Fuzzy Merging Algorithm

Advisor: Dr. Carl Looney

Committee member: Dr. Yaakov Varol
Committee member: Dr. Shunfeng Song

Liyan Zhang
Computer Science Department
University of Nevada, Reno
Reno, NV 89557
[email protected]

May, 2001

0
 Comparison of Fuzzy c-means Algorithm and New Fuzzy
Clustering and Fuzzy Merging Algorithm

Abstract: Clustering is the process of grouping feature vectors into classes in

the self-organizing mode. Choosing cluster centers is crucial to the clustering.
In this paper we compared two fuzzy algorithms: fuzzy c-means algorithm and
the new fuzzy clustering and fuzzy merging algorithm. Fuzzy c-means algorithm
uses the reciprocal of distances to decide the cluster centers. The representation
reflects the distance of a feature vector from the cluster center but does not
differentiate the distribution of the clusters. The new fuzzy clustering and
merging algorithm uses Gaussian weights and the generated cluster centers are
more representative. When a feature vector is of equal distance from two cluster
centers, it weighs more on the widely distributed cluster than on the centrally
located cluster. We implemented both algorithms in Java and tested them on
several data sets.

Keywords: fuzzy c-means, fuzzy clustering, classification, pattern recognition.

1. Introduction
Clustering is the process of grouping feature vectors into classes in the self-
organizing mode. Let {x(q): q = 1,…,Q} be a set of Q feature vectors. Each feature vector
x(q) = (x1(q), …, xN(q)) has N components. The process of clustering is to assign the Q
feature vectors into K clusters {c(k): k = 1, …, K} usually by the minimum distance
assignment principle.

Choosing the representation of cluster centers (or prototypes) is crucial to the

clustering. Feature vectors that are farther away from the cluster center should not have
as much weight as those that are close. These more distant feature vectors are outliers
usually caused by errors in one or more measurements or a deviation in the processes that
formed the object [9].

The simplest weighting method is arithmetic averaging. It adds all feature vectors in
a cluster and takes the average as prototype. Because of its simplicity, it is still widely
used in the clustering initialization.

The arithmetic averaging gives the central located feature vectors the same weights
as outliers. To lower the influence of the outliers, median vectors are used in some
proposed algorithms.

1
 To be more immune to outliers and more representative, the fuzzy weighted average
is introduced to represent prototypes:

Zn (k) = {q: q k} wqkx(q)n; (1)

Rather than a Boolean value 1 (true, which means it belongs to the cluster) or 0 (false,
does not belong), the weight wqk in equation (1) represent partial membership to a
cluster. It is called a fuzzy weight. There are different means to generate fuzzy
weights.

One way of generating fuzzy weights is the reciprocal of distance

wqk = 1/ Dqk , (wqk= 1 if Dqk = 0) (2)

It was used in earlier fuzzy clustering algorithms [2]. When the distance between the
feature vector and the prototype is large, the weight is small. On the other hand, it is large
when the distance is small.

Using Gaussian functions to generate fuzzy weights is the most natural way for
clustering. It is not only immune to outliers but also provides appropriate weighting for
more centrally and densely located vectors. It is used in the fuzzy clustering and fuzzy
merging (FCFM) algorithm.

In this project, we implemented the fuzzy c-means (FCM) algorithm and the fuzzy
clustering and merging algorithm in Java, applied the algorithms to several data sets and
compared the weights of the two algorithms.

Section 2 describes the background of clustering algorithms. Section 3 discusses the

implementation of FCM algorithm and the results on data sets. Section 4 presents the
results using the FCFM algorithm. Section 5 compares the fuzzy weights in FCM and the
Gaussian weights in the FCFM algorithm.

2. Clustering Algorithms
The clustering groups a sample set of feature vectors into K clusters via an appropriate
similarity (or dissimilarity) criterion (such as distance from the center of the cluster).

2.1 The k-means Algorithm

The k-means algorithm assigns feature vectors to clusters by the minimum distance
assignment principle [5], which assigns a new feature vector x(q) to the cluster c(k) such
that the distance from x(q) to the center of c(k) is the minimum over all K clusters. The
basic k-means algorithm is as follows:
 Put the first K feature vectors as initial centers
 Assign each sample vector to the cluster with minimum distance assignment
principle.

2
 Compute new average as new center for each cluster
 If any center has changed, then go to step 2, else terminate.

The advantages of the method are its simplicity, efficiency, and self-organization. It
is used as initial process in many other algorithms. The disadvantages are: 1) K must be
provided; 2) it is a linearly separating algorithm.

2.2 The ISODATA Algorithm

This algorithm is based on the k-means algorithm, and employs processes of
eliminating, splitting, and clustering. The algorithm is described as following [8].
 Start with Kinit (initial number of clusters) which is user-given. Assign the first Kinit
samples as cluster centers.
 Assign all samples to the clusters by minimum distance principle.
 Eliminate clusters that contain less than nmin feature vectors and reassign those vectors
to other clusters to yield K clusters.
 Compute a new cluster center as the average of all feature vectors in each cluster.
 For each kth cluster, compute the mean-squared error n2(k) of each nth component xn
over that cluster and find the maximum n*2(k) component mean-squared error over
within cluster k for over n = 1, …, N, where the index n* is for the maximum
component.
 If there are not enough clusters (K init < K/2) and this is not the last iteration, then if
max(k) > split for any cluster k, split that cluster into two.
 If this is an even iteration and Kinit > 2K, then compute all distances between cluster
centers. Merge the clusters that are close than a given value.

The advantages of the ISODATA are its self-organizing capability, its flexibility in
eliminating clusters that are too small, its ability to divide clusters that are too dissimilar,
and its ability to merge clusters that are sufficiently similar. Some disadvantages are: 1)
multiple parameters must be given by the user, although they are not known a priori; 2) a
considerable amount of experimentation may be required to get reasonable values; 3) the
clusters are ball shaped as determined by the distance function; 4) the value determined
for K depends on the parameters given by the user and is not necessarily the best value;
and 5) a cluster average is often not the best prototype for a cluster [9].

2.3 Fuzzy Clustering Algorithms

Fuzzy clustering plays an important role in solving problems in the areas of pattern
recognition and fuzzy model identification. A variety of fuzzy clustering methods have
been proposed and most of them are based upon distance criteria [6]. One widely used
algorithm is the fuzzy c-means (FCM) algorithm. It uses reciprocal distance to compute
fuzzy weights. A more efficient algorithm is the new FCFM. It computes the cluster
center using Gaussian weights, uses large initial prototypes, and adds processes of
eliminating, clustering and merging. In the following sections we discuss and compare
the FCM algorithm and FCFM algorithm.

3
3. The Fuzzy c-means Algorithm
The fuzzy c-means (FCM) algorithm was introduced by J. C. Bezdek [2]. The idea of
FCM is using the weights that minimize the total weighted mean-square error:

J(wqk, z(k)) =  (k=1,K)  (k=1,K) (wqk)|| x(q)- z(k)||2 (3)

 (k=1,K) (wqk) = 1 for each q

wqk = (1/(Dqk)2)1/(p-1) /  (k=1,K) (1/(Dqk)2)1/(p-1) , p > 1 (4)

The FCM allows each feature vector to belong to every cluster with a fuzzy truth
value (between 0 and 1), which is computed using Equation (4). The algorithm assigns a
feature vector to a cluster according to the maximum weight of the feature vector over all
clusters.

3.1 The FCM Algorithm

Based on the original FCM algorithm [9], we added Steps 7, 8, 9 and modified Step
2. The pseudocode of our FCM algorithm follows.
// K is initial number of clusters, Imax is the iteration of fuzzy
// c-means, p is for the weight
Input initial number of clusters K, Imax, p

------------step 1: --------------
//initialize weights of prototype
for k = 0 to K-1
for q = 0 to Q-1
w[q,k] = random();

------------step 2: --------------
//standardize the initial weight over K
for q = 0 to Q-1
sum = 0.0;
for k = 0 to K-1
sum = sum + w[q,k];
for k = 0 to K-1
w[q,k] = w[q,k] /sum;

*****************************************
// starting fuzzy c-means loop
I = 0

------------step 3: --------------
// standardize cluster weights over Q
for k = 0 to K-1
min = 99999.0; max =0.0;
for q = 0 to Q-1

4
 if (w[q,k] > max)
max = w[q,k];
if (w[q,k] < min)
min = w[q,k];

sum = 0.0
for q = 0 to Q-1
sum = sum + (w[q,k] – min) /( max –min);

for q = 0 to Q-1
w[q,k] = w[q,k]/sum;

------------step 4: --------------
// compute new prototype center
for k = 0 to K-1
for n = 0 to N-1
sum = 0.0;
for q = 0 to Q-1
sum = sum + w[q,k] x[n,q];
z[n,k] = sum;

------------step 5: --------------
// compute new weight
for q = 0 to Q-1
sum = 0.0
for k = 0 to K-1
D[q,k] =0.0;
for n = 0 to N-1
D[q,k] = D[q,k] + (x[n,q] – z[n,k])2
sum = sum + (1/(1 + D[q,k]))1/(p-1) ;
for k = 0 to K-1
W[q,k] = (1/(1 + D[q,k]))1/(p-1) /sum;

------------step 6: --------------
I = I + 1

If I < Imax
Goto step 3;
// end of fuzzy c-means loop
**********************************************

------------step 7: --------------
// assign feature vector according the max weight
for q = 0 to Q-1
maxWeight = 0.0;
for k = 0 to K-1
if maxWeight < weight[q,k];
maxWeight = weight[q,k];
kmax = k;
cluster[q] = k;

------------step 8: --------------
// eliminate clusters with no feature vectors
eliminate(0); /* call the process of eliminating clusters contains
less than or equal to the number passed to it.
Here we only pass 0 for this algorithm. */

5
 ------------step 9: --------------
// compute arithmetic center of clusters
// calculate sigma and Xie_Beni value
.

for k = 1 to K do
fuzzyweights(); /* Calculate fuzzy weight (Eqn. 4)
2 = variance();
k
/* Get variance (mean-square error) of
each cluster (Eqn. 9) */
 = Xie-Beni(); /* Compute modified XB (Eqn. 8) */

3.2 A New Fuzzy c-means Implementation

3.2.1 Initialize the Fuzzy Weights. In order to comparing the FCM with FCFM, our
implementation allows the user to choose initializing the weights using feature vectors or
randomly. The process of initializing the weights using feature vectors assigns the first
Kinit (user-given) feature vectors to prototypes then computes the weights by Equation (4).

3.2.2 Standardize the Weights over Q. During the FCM iteration, the computed cluster
centers get closer and closer. To avoid the rapid convergence and always grouping into
one cluster, we use

w[q,k] = (w[q,k] – wmin)/( wmax – wmin) (5)

before standardizing the weights over Q. Where wmax, wmin are maximum or minimum
weights over the weights of all feature vectors for the particular class prototype.

3.2.3 Eliminating Empty Clusters. After the fuzzy clustering loop we add a step (Step
8) to eliminate the empty clusters. This step is put outside the fuzzy clustering loop and
before calculation of modified XB validity. Without the elimination, the minimum
distance of prototype pair used in Equation (8) may be the distance of empty cluster pair.
We call the method of eliminating small clusters by passing 0 to the process so it will
only eliminate the empty clusters.

3.2.4 Modified XB. After the fuzzy c-means iteration, for the purpose of comparison and
to pick the optimal result, we add Step 9 to calculate the cluster centers and the modified
Xie-Beni clustering validity  [7]:

The Xie-Beni validity is a product of compactness and separation measures [10]. The
compactness-to-separation ratio  is defined by Equation (6).

 = {(1/K)(k=1,K) k2}/Dmin2 (6)

k2 = (q=1,Q) wqk || x(q) – c(k) ||2 (7)

Dmin is the minimum distance between the cluster centers.

6
The Modified Xie-Beni validity  is defined as

 = Dmin2/ {(k=1,K) k2 } (8)

The variance of each cluster is calculated by summing over only the members of
each cluster rather than over all Q for each cluster, which contrasts with the original Xie-
Beni validity measure.

k2 = {q: q is in cluster k} wqk || x(q) – c(k) ||2 (9)

3.3 The Fuzzy c-means Results

We used the following data sets to run both algorithms:
The iris data set [7] is Anderson’s [1] 150 feature vectors of iris species. The data is
labeled as K = 3 that represents 3 subspecies (Sestosa, Versicolor and Virginica). Each
feature vector has 4 features. The given sample contains 50 labeled feature vectors from
each class.
The test13 data set consists of 2 natural classes. There are 13 two-dimensional
feature vectors. Figure 1 shows the feature vectors.
The Wisconsin breast cancer data set (WCBD) [7] consists of 200 randomly selected
from more than 500 breast cancer vectors of the University of Wisconsin Medical School.
Each feature vector has 30 features in [0, 1]. The vectors are labeled for two classes. One
label is attached to 121 vectors while the other is attached to 79 vectors.
The geological data set [7] is labeled for K = 2 classes. Each of the Q = 70 feature
vectors has N = 4 features. The data labels are estimates by humans that give 35 to each
class. These were assigned by humans providing their best guesses.

3.3.1 Different Combinations of Parameters. We tried different combinations of

parameters in Table 1 and Table 2. The different combinations of Kinit, iterations and p
(Equation (4)) result in different outputs.

In Table 1, 2, 3 and 4, we run the FCM without eliminating empty clusters. That is, if
iteration I = 300, we did not interrupt the process until it reaches to the 300 iterations,
then we deleted the empty clusters and then compute the modified XB.

Table 1 presents the results on iris data set. We standardized the components of
feature vectors between 0 and 1, set K init = 150, assigned the feature vectors to the
prototypes, and varied the value p and iteration number, p = 2, 3, 4, 5 and iteration = 100,
200, 300. The best result we get in Table 1 is when p = 2 and iteration >= 200, the
modified XB = 2.365, there are two clusters, one has 56 feature vectors, the other has 94.

7
 Table 1. Results on the iris date set via FCM algorithm (Kinit = 150).
Iterations P=2 P=3 P=4 P=5
Clusters XB Clusters XB Clusters XB Clusters XB
100 56, 57, 37 2.19E-30 48, 93, 9 1.8E-31 53, 68, 25, 4 0.39E-32 58, 92 2.175
200 56, 94 2.365 57, 93 2.225 57, 93 2.190 58, 92 2.175
300 56, 94 2.365 57, 93 2.225 57, 93 2.190 58, 92 2.175

Table 2 shows the results on iris data set under the same condition as in Table 1 except
Kinit = 50. The best result so far was when p = 2 and iteration I >= 100, the modified XB
= 2.385, there are two clusters, one has 56 feature vectors, the other has 94.

Table 2. Results on the Iris Data via FCM Algorithm (Kinit = 50).
Iterations P=2 P=3 P=4 P=5
Clusters XB Clusters XB Clusters XB Clusters XB
100 56, 94 2.385 51, 93, 6 9.46E-32 53, 93, 4 1.36E-31 52, 92, 3, 3 3.35E-32
200 56, 94 2.385 57, 93 2.225 57, 93 2.190 52, 92, 6 1.10E-31
300 56, 94 2.385 57, 93 2.225 57, 93 2.190 54, 92, 4 2.00E-31

3.3.2 Initializing the Prototypes. To study the difference between initializing prototypes
randomly and using the feature vectors, we run the program at same conditions but
initialized the prototypes using feature vectors in Table 3 and initialized randomly in
Table 4. The resulting clusters were affected by the initial prototype centers.

Table 3 shows the results on iris data set. We standardized the feature vectors
into[0,1], initialized prototyes using the first Kinit feature vectors, and run the program by
fixing p = 2 and varying the iteration number of fuzzy clustering and K init, where iteration
number I = 100 and 200, Kinit = 150, 120, 90, 60, 30, 10 and 5. The results were very
similar, after 200 iteration, the first cluster contains 56 feature vectors and the second
contains 94 feature vectors.

Table 3. Results on Iris Data via the FCM Algorithm When Prototypes Are
Initialized Using Feature Vectors.

Kinit I = 100 I = 200

Clusters XB Clusters XB
150 56, 57, 37 2.19E-30 56, 94 2.365
120 56, 94 2.355 56, 94 2.355
90 56, 94 2.355 56, 94 2.355
60 56, 94 2.355 56, 94 2.355
30 56, 94 2.405 56, 94 2.405
10 56, 94 2.355 56, 94 2.355
5 50, 94, 6 1.15E-31 56, 94 2.355

8
 Table 4 presents the results of the FCM algorithm under the same conditions as in
Table 3 except initializing the prototypes randomly.

Table 4. Results on Iris Data via the FCM Algorithm When Prototypes Are
Initialized Randomly.
Kinit I = 100 I = 200 I = 300
Clusters XB Clusters XB Clusters XB
150 59, 91 2.175 59, 91 2.175 59, 91 2.175
120 54, 90, 3, 3 3.05E-32 59, 90, 1 3.63E-31 60, 90 2.12
90 60, 90 2.12 60, 90 2.17 60, 90 2.12
60 59, 91 2.17 59, 91 2.17 59, 91 2.17
30 20, 90, 2, 7, 31 2.08E-32 52, 90, 8 5.07E-31 60, 90 2.105
10 58, 73, 19 2.87E-31 58, 92 2.195 58, 92 2.195
5 56, 94 2.355 56, 94 2.355 56, 94 2.355

3.3.3 Over Convergence. Figure 1 shows the feature vectors of Test13. There are two
class. We run the FCM algorithm on Test13. We standardized the feature vector,
initialized the prototypes using feature vectors, and set the Kinit = 13. After 30 iterations,
yielded two clusters, K = 2. We added 10 iterations, K became 1. The result shows in
Figure 2.

Figure 1. Feature Vectors on Test13 Figure 2. Clustering Results on Test13

K = 2. K = 1.

Feature Vectors on Test13 FCM Result on Test13

3.5 3.5

3 3

2.5 2.5

2 2
y

1.5 1.5

1 1

0.5 0.5

0 0
0 0.2 0.4 0.6 0 0.2 0.4 0.6
x x

9
3.4 More Results

Table 5 represents the results on WBCD data set. The original prototype number K init
was 40, p = 3, the prototypes were initialized using feature vectors. We run fuzzy c-
means by gradually adding iterations as shown in column 1. Between iterations, the
empty prototypes are deleted. After the total iteration 569, the first cluster does not
change any more, the centers of the last two get very close. As some feature vectors wave
between the two clusters, they are hard to merge to one.

Table 5. Results on WBCD Data via the FCM Algorithm.

Total Cluster Clusters XB
Iterations Iterations Number
200 200 23 26, 14, 18, 4, 3, 2, 3, 71, 1, 1, 6.83E-34
13, 7, 3, 3, 8, 2, 1, 1, 14, 1, 1,
2, 1
+ 300 500 10 33, 3, 6, 2, 20, 1, 4, 124, 6, 1 1.95E-32
+ 30 530 3 52, 25, 123 5.60E-10
+2 532 3 69, 62, 69 4.43E-11
+2 534 3 72, 56, 72 7.0E-13
+5 539 3 72, 53, 75 9.3E-17
+ 30 569 3 72, 126, 2 1.25E-32
+1 570 3 72, 113, 15 9.43E-32
+1 571 3 72, 113, 15 1.88E-31
+ 1000 1571 3 72, 119, 9 3.96E-32
+ 20 1591 3 72, 115, 72 6.73E-32
+ 1000 2591 3 72, 120, 8 2.45E-32
+ 5000 7591 3 72, 106, 26 6.13E-32
+ 1000 8591 3 72, 111, 17 1.38E-31
+ 10000 18591 3 72, 94, 34 6.7E-32

When we tried to set the iteration 18591 without eliminating the empty clusters
during the loop, the results are very different. We only eliminated the empty clusters until
out of the fuzzy clustering iteration and before the calculation of modified XB validity,
The result is K = 10 and the count of each cluster is 26, 15, 3, 22, 2, 1, 2, 127, 1 and 1.
The modified XB is 1.64E-31.

Table 6 shows the results on geological data using the FCM algorithm. The initial
number of prototype Kinit was 70, p = 3. The prototypes were initialized using feature
vectors. We started with 100 iterations, then added 10 each time until K=2 is reached.

10
 Table 6. Results on Geological Data via the FCM Algorithm.
Total Cluster Clusters XB
Iterations Iterations Number
100 100 4 20, 1, 29, 20 1.83E-32
+ 10 110 3 21, 29, 20 2.00E-29
+ 10 120 3 21, 40, 9 6.5E-32
+ 10 130 3 21, 30, 19 1.65E-31
+ 10 140 3 21, 48, 1 2.21E-32
+ 10 150 2 21, 49 0.543

4. The New Fuzzy Clustering and Fuzzy Merging (FCFM) Algorithm

4.1 The FCFM Algorithm

The FCFM algorithm [7] uses Gaussian weighted feature vectors to represent the
prototypes. The algorithm starts with large numbers of clusters and eliminates clusters by
distance or by size so the prototypes are much more representative. The FCFM algorithm
computes fuzzy weights using Equation (10)

p(r) = exp[-(xp - (r) )2/(22)(r)]/{ (m = 1,P) exp[-(xm - (r) )2/(22)(r)]} (10)

The FCFM algorithm does the following steps:

 Standardize the Q sample feature vectors and use a large Kinit.
 Eliminate the prototypes that are closer to other prototype than a distance threshold
DThresh, which can be set by a user.
 Apply k-means as the first step to get the prototypes.
 Eliminate small clusters.
 Loop in computing the fuzzy weights and MWFEV (modified weighted fuzzy
expected value) for each cluster to obtain the prototype and then assign all of the
feature vectors to clusters based on the minimum distance assignment. End the loop
when the fuzzy centers do not change.
 Merge clusters.

4.1.1 Initial K. The FCFM algorithm uses a relatively large Kinit to thin out the
prototypes. The default Kinit is calculated as:

Kinit = max{6N + 12log2Q, Q} (11)

4.1.2 Modified Xie-Beni Validity. The FCFM uses modified Xie-Beni validity same as
in FCM (3.2.4).

4.1.3 Clustering. To obtain a more typical vector to represent a cluster, the algorithm
uses modified weighted fuzzy expected value as the prototypical value:

11
  (r+1) = {p=1, P} p(r) x p (12)

 (r+1) is obtained by Picard iterations. The initial value (0) is the arithmetic average of
the set of real vectors.

p(r) = exp[-(xp - (r) )2/(22)(r)]/{ (m=1,P) exp[-(xm - (r) )2/(22)(r)]} (13)

2 is the mean-square error.

The process computes the fuzzy weights and the MWFEV componentwisely for each
cluster to obtain the cluster center. It also computes the mean-square error 2 for each
cluster, and then assigns each feature vector to a cluster by the minimum distance
assignment principle.

4.1.4 Merging. For every pair of prototypes, the algorithm computes the distance
between them. It finds the pair with shortest distance. If the pair meets the merge criteria,
the two clusters are merged into one.

4.2 The Results of the FCFM Algorithm

We run the FCFM on three data sets, which are iris data set, geological data and
WBCD data set. The results are presented below.

For iris data, the initial number of prototypes was set to 150. The cluster centers were
initialized randomly. After eliminating clusters by distance (0.28), K was reduced to 58.
Reduce K to 21, 20, 18, 16, 14, 12 and 10 respectively by eliminating small clusters with
p = 1, 2, 3, 4, 5, 6 and 8. Then set iteration number of fuzzy clustering to 100. Merge with
 = 0.5, 0.52, 0.54, 0.56 and 0.6, the K for each merge was 7, 4, 3, 3 and 2. Some more
detailed results are in Table 7.

Table 7. Result of the FCFM Algorithm on Iris Data.

K Clusters XB k
4 49, 34, 17, 50 0.825 0.204, 0.227, 0.134, 0.192
3 63, 37, 50 1.45 0.233, 0.228, 0.192
3 62, 38, 50 1.4 0.231, 0.229, 0.192
2 100, 50 5.82 0.323, 0.192

Table 8 shows the results on WBCD data set. The original prototypes were set to
200 and initialized using feature vectors. We started at deletion of close prototypes (d =
0.83), the prototype became 29. We then reduced the K to 14, 13, 11, 9 by eliminating
small clusters with p = 1, 2, 4 and 7. Set fuzzy iteration f = 40. We reduced the K to 7, 5,
4, 3, 2 by selecting  = 0.5, 0.6, 0.66, 0.7 and 0.8 in merge process.

12
 Table 8. Results on WBCD Data via the FCFM Algorithm.
K Clusters XB k
5 20, 18, 42, 14, 106 0.234 0.59, 0.83, 0.62, 0.77, 0.66
4 29, 18, 126, 27 0.292 0.59, 0.82, 0.72, 0.82
3 44, 123, 33 0.42 44, 123, 33
2 125, 75 1.082 75, 125

Table 9 presents the results on geological data set. The original prototypes were
initialized using feature vectors. After deletion of close prototypes (d = 0.34), K reduced
to 23. We eliminated small clusters with p = 1, 2, 3, 4 and 5, K became to 14, 11, 9, 8 and
7. The fuzzy clustering iteration was set to 100. Merge the clusters with  = 0.5, 0.55,
0.65, 0.85, 0.9, K reduced to 6, 5, 4, 3 and 2.

Table 9. Results on Geological Data via the FCFM Algorithms

K Clusters XB k
5 14, 8, 27, 12, 9 0.612 0.44, 0.32, 0.30, 0.40, 0.32
4 14, 10, 34, 12 0.995 0.42, 0.32, 0.45, 0.40
3 16, 21, 33 0.87 0.42, 0.60, 0.41
3 17, 53 2.39 0.47, 0.58

5. Comparing Weights of the FCM and the FCFM Algorithm

5.1 The Weight of Fuzzy c-means

In the FCM algorithm, the weights are defined in Equation (4). From the figure
3, we can see that the weight of feature vectors to a prototype is not related to how
the feature vectors are distributed in the cluster. When the p value increases, the
weight difference between two feature vectors decreases, so the speed of
convergence increases.

Only the distance between the feature vector and the center decides the weight of the
vector. When the clustering continues, two centers of not well-separated clusters become
closer and closer until they finally merge into a large one.

The prototype representation does not reflect the distribution of feature vectors in the
cluster. The feature vector with equal distance to two clusters will weigh the same, no
matter that the clusters are centrally located or evenly distributed.

13
 Figure 3. Weights of the FCM Algorithm For Different p Values

Weights of Fuzzy c-means, p=2 Weights of Fuzzy c-means, p=3

1.2 1.2

1 1

0.8 0.8

f(x)
0.6
f(x)

0.6

0.4 0.4

0.2 0.2

0 0
-2 -1 0 1 2 -2 -1 0 1 2

x x

Weighs of Fuzzy c-mean, p=4 Weights of fuzzy c-means, p=5

1.2 1.2

1 1

0.8 0.8
f(x)

f(x)

0.6 0.6

0.4 0.4

0.2 0.2

0 0
-2 -1 0 1 2 -2 -1 0 1 2
x x

5.2 The Weights in the FCFM (Gaussian Weights)

In the FCFM algorithm, the weights are defined by Equation (13). Figure 4 shows
that the weight of Gaussian is decided not only by the distance between the feature vector
and the prototype, but also the distribution of the feature vectors in the cluster. When the
feature vectors in a cluster are centrally and densely distributed around the center, the

14
sigma value will be small. The features that are close to prototype weigh much more than
the farther features. The weight of a feature vector is related to the shape of the Gaussian.
If a feature vector has equal distance from two prototypes, it weighs more on the more
widely distributed cluster than on the centrally located cluster. Thus, Gaussian fuzzy
weights are more immune to outliers and more representative than the other kind of fuzzy
weights.

Figure 4. Weights of Gaussian for Different Sigma Values.

Weights of Gaussian, sigma = 0.1 Weight of Gaussian, sigma=0.2

1.2 1.2

1 1

0.8 0.8

0.6

f(x)
0.6
f(x)

0.4
0.4
0.2
0.2
0
-1.5 -1 -0.5 0 0.5 1 1.5 0
-0.2 -1.5 -1 -0.5 0 0.5 1 1.5
x x

Weights of Gaussian, sigma=0.3 Weights of Gaussian, sigma=0.4

1.2
1.2

1
1

0.8
0.8
f(x)

0.6
f(x)

0.6

0.4
0.4

0.2
0.2

0
0
-1.5 -1 -0.5 0 0.5 1 1.5
-1.5 -1 -0.5 0 0.5 1 1.5
x
x

15
 Weights of Gaussian, sigma=0.5

1.2

0.8
f(x)

0.6

0.4

0.2

0
-1.5 -1 -0.5 0 0.5 1 1.5
x

6. Implementation of the FCM and FCFM in Java

A Java GUI application, which implements both the FCM algorithm and the FCFM
algorithm, is developed (see Appendix 1). The application consists of following classes:
 Features: It stores the feature vectors into a two-dimensional array, and contains
methods for standardizing the feature vectors. It is used as a member in
Prototypes.
 Prototypes: It stores the prototypes, count of cluster and implements methods for
assigning features into clusters, computing modified XB validity, k-means etc. It is
the base class of FCmentPrototype and FClustPrototype.
 FClustPrototype: This class inherits Prototypes and implements the FCFM
algorithm.
 FCmeanPrototype: It inherits Prototypes, and has its own methods of
initializing feature vectors, standardizing weights, and calculating of cluster centers.
The class implements the FCM algorithm.
 FCluster: It is the driver of the two algorithms. It interacts with user and calls the
methods in FCmentPrototype and FclustPrototype to run the FCM and the
FCFM.
 FormatWriter: It is a utility class for formatting and outputting results.
 FormattedInput: It is a utility class for getting user input.

16
 Using this console application, a user can choose the data file, decide whether
standardizing the feature vectors, set initial number of clusters, and the number of
clustering iterations. The results are written to a file.

7. Conclusions and Future Work

The FCM algorithm uses reciprocal distance to compute the fuzzy weights. When a
feature vector is of equal distance from two cluster centers, it weights the same on the
two clusters no matter what is the distribution of the clusters. It cannot differentiate the
two clusters with different distributions of feature vectors. Therefore, the FCM algorithm
is more suited to data that is more or less evenly distributed around the cluster centers.
The FCM algorithm lumps the two clusters with natural shapes but close boundaries into
a large cluster. For some difficult data such as WBCD data, it is hard to for the FCM to
cluster the very closed classes together without the help of other mechanisms such as
elimination of small clusters.

The FCFM algorithm uses Gaussian weights, which are most representative and
immune to outliers. Gaussian weights reflect the distribution of the feature vectors in the
clusters. For a feature vector with equal distance from two prototypes, it weighs more on
the widely distributed cluster than on the narrowly distributed cluster. The FCFM
algorithm outperforms the FCM on all the test data we used in this paper.

Since Java language is tightly integrated with the Internet, we may further develop a
Java Applet using the four main classes, Feature, Prototypes,
FClustPrototype, and FcmeanPrototype. By embedding the applet in a Web
page, users can easily access and use the FCM and the FCFM algorithm through
browsers. This is a candidate for future work. The code can be obtained at
http://www.cs.unr.edu/~lzhang/fuzzyCluster/paper/fcgui/FCluster.jar.

8. Acknowledgment
I would like to express my sincere appreciation to Dr. Carl Looney for guiding me
through the project, for his thoughtful advice, insight and detailed suggestions. Many
thanks and appreciation go to Dr. Yaakov Varol and Dr. Shunfeng Song for serving as
my graduate committee members and previewing the paper. I thank Dr. Kenneth
McGwire for giving me the chance working at Desert Research Institute, we not only
learned knowledge from working on the project but also learned a lot from his great
personality. I thank Luisa Bello for helping me edit the paper.

9. References
[1] E. Anderson, “The iris of the Gaspe peninsula” Bulletin American Iris Society, Vol.
59, 2-5, 1935.
[2] J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum
Press, New York 1981.
[3] J. C. Bezdek, etc. Convergence Theory for Fuzzy c-Means: Counterexamples and
Repairs, IEEE Trans. Syst., September/October 1987.

17
[4] Maria Colmenares & Olaf WolkenHauer, “An Introduction into Fuzzy Clustering”,
http://www.csc.umist.ac.uk/computing/clustering.htm, July 1998, last update 03 July,
2000
[5] Marti Hearst, K-Means Clustering, UCB SIMS, Fall 1998,
http://www.sims.berkeley.edu/courses/is296a-3/f98/lectures/ui-bakground/sld025.htm.
[6]Uri Kroszynski and Jianjun Zhou, Fuzzy Clustering Principles, Methods and
Examples, IKS, December 1998
[7] Carl G. Looney A Fuzzy Clustering and Fuzzy Merging Algorithm, CS791q Class
Notes, http://www.cs.unr.edu/~looney/.
[8] Carl G. Looney Pattern Recognition Using Neural Networks, Oxford University
Press, N.Y., 1997.
[9] Carl G. Looney “Chapter 5. Fuzzy Clustering and Merging”, CS791q Class Notes,
http://www.cs.unr.edu/~looney/.
[10]Ramze Rezaee M, Lelieveldt B P F and Reiber J H C, A New Cluster Validity Index
for the Fuzzy c-mean, Pattern Recognition Letters, (Netherlands) Mar 1998.
[11] M.J.Sabin, Convergence and Consistency of Fuzzy c-means /ISODATA Algorithms,
IEEE Trans. Pattern Anal. Machine Intell., September 1987.

18
Appendix 1. The Java Source Code For the Paper

The application has following classes:

 FuzzyApp: It contains main() method of the application.

 FuzzyFrame: It is the interface of the two algorithms. It interacts with user and
calls the methods in FCmentPrototype and FclustPrototype to run the
FCM and the FCFM.
 Features: It stores the feature vectors into a two-dimensional array, and contains
methods for standardizing the feature vectors. It is used as a member in
Prototypes.
 Prototypes: It stores the prototypes, count of cluster and implements methods for
assigning features into clusters, computing modified XB validity, k-means etc. It is
the base class of FCmentPrototype and FClustPrototype.
 FClustPrototype: This class inherits Prototypes and implements the FCFM
algorithm.
 FCmeanPrototype: It inherits Prototypes, and has its own methods of
initializing feature vectors, standardizing weights, and calculating of cluster centers.
The class implements the FCM algorithm.
 FormatWriter: It is a utility class for formatting and outputting results.
 FormattedInput: It is a utility class for getting user input.
 FuzzyFrame_AboutBox: It is the dialog for the about button.

These classes can be contained in FCluster.jar at the directory

http://www.unr.edu/~lzhang/fuzzyCluster/paper/fcgui.
To open the jar file:
jar xvf FCluster.jar
To compile the source code:
javac FuzzyApp.java
To run the program:
java FuzzyApp

The test13.dta is an example of the data input file. The first line is the number of
components N. The second line K is the number of initial classes. The third line J is the
number of output components. The forth line Q is the number of vectors. In this two
algorithms, J and K are not in use.

19