a robust fuzzy classification maximum likelihood clustering...
TRANSCRIPT
International Journal of Uncertainty,
Fuzziness and Knowledge-Based Systems
Vol. 21, No. 5 (2013) 755−776
© World Scientific Publishing Company
DOI: 10.1142/S0218488513500360
755
A ROBUST FUZZY CLASSIFICATION MAXIMUM
LIKELIHOOD CLUSTERING FRAMEWORK
MIIN-SHEN YANG*,‡, CHIH-YING LIN* and YI-CHENG TIAN*,†
*Department of Applied Mathematics
Chung Yuan Christian University
Chung-Li 32023, Taiwan
Department of Applied Mathematics, †Center for General Education,
Hsin Sheng College of Medical Care and Management,
Chung-Li, Taiwan ‡[email protected]
Received 16 October 2012
Revised 26 April 2013
In 1993, Yang first extended the classification maximum likelihood (CML) to a so-called fuzzy
CML, by combining fuzzy c-partitions with the CML function. Fuzzy c-partitions are generally an
extension of hard c-partitions. It was claimed that this was more robust. However, the fuzzy CML
still lacks some robustness as a clustering algorithm, such as its in-ability to detect different volumes
of clusters, its heavy dependence on parameter initializations and the necessity to provide an a priori
cluster number. In this paper, we construct a robust fuzzy CML clustering framework that has a
robust clustering method. The eigenvalue decomposition of a covariance matrix is firstly considered
using the fuzzy CML model. The Bayesian information criterion (BIC) is then used for model
selection, in order to choose the best model with the optimal number of clusters. Therefore, the
proposed robust fuzzy CML clustering framework exhibits clustering characteristics that are robust
in terms of the parameter initialization, robust in terms of the cluster number and also in terms of its
capability to detect different volumes of clusters. Numerical examples and real data applications
with comparisons are provided, which demonstrate the effectiveness and superiority of the proposed
method.
Keywords: Clustering; model-based clustering; k-means; fuzzy clustering; fuzzy c-means;
classification maximum likelihood; robustness.
1. Introduction
Data analysis is the science of analyzing real world data. Cluster analysis is a useful tool
for data analysis. It identifies clusters as the groups of data points in a data set with the
greatest similarity within the same cluster and the largest dissimilarity between different
clusters. Cluster analysis is a branch of statistical multivariate analysis.1,2
In contrast,
learning and recognition mostly begin with clustering, so cluster analysis represents an
unsupervised learning in pattern recognition.3.4
From the statistical point of view,
*Corresponding author.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 756
clustering methods can be generally categorized as either having a (probability) model-
based approach or a nonparametric approach. A model-based clustering approach
assumes that the data set obeys a mixture model of probability distributions, so a mixture
likelihood approach to clustering is used.5,6
For a nonparametric approach, clustering
methods are generally based on an objective function of similarity or dissimilarity
measures and partitional methods are popularly used.2,4
Most partitional methods suppose
that the data set can be represented by finite cluster prototypes with their own objective
functions. Therefore, the definition of the dissimilarity (or distance) between points and
cluster prototypes is essential for partitional methods. The most popular partitional
methods for cluster prototypes are k-means,7,8
trimmed k-means,9,10
fuzzy c-means
(FCM),11−13
mean shift14−16
and possibilistic c-means (PCM).17−19
There are two categories of model-based clustering algorithms. One is based on the
expectation and maximization (EM) algorithm,20,21
and the other is based on the
classification maximum likelihood (CML). 22−24
The EM algorithm is most popular, but it
is sensitive to initial values and to cluster shapes in which an a priori number of
components (clusters) must be given.21
Scott and Symons23
first proposed the CML
procedure. Yang25
then extended the original CML to the so-called fuzzy CML (FCML)
based on fuzzy c-partitions. Banfield and Raftery26
extended the original CML by using
the eigenvalue decomposition of a covariance matrix to allow the specification of which
feature (orientation, size and shape) is common to all clusters and which differs between
clusters. Fraley and Raftery27,28
continued to develop a more general model-based
clustering procedure.
In the literature, there are some applications of the fuzzy CML.29-33
However, the
FCML is not sufficiently robust in clustering, so it cannot detect different volumes (sizes
and shapes) of clusters and also requires an a priori number of clusters with parameter
initialization. In this paper, a robust framework is constructed for FCML clustering. It is
firstly necessary to consider the eigenvalue decomposition of a covariance matrix in the
fuzzy CML model. The Bayesian information criterion (BIC) is then used to choose the
best model with the optimal number of clusters. The remainder of this paper is organized
as follows. Section 2 provides a review of the FCML clustering method proposed by
Yang.25
In Sec. 3, the robust FCML clustering framework is constructed. Section 4
contains some numerical comparisons of the proposed method with existing methods.
The method is then applied to real data sets. Finally, conclusions are stated in Sec. 5.
2. Fuzzy Classification Maximum Likelihood Clustering
The classification maximum likelihood (CML)22−24
is a classification approach that
approximates the maximum likelihood estimates. Let ,..., 1 nxxX = be a random
sample of size, n , from a population that consists of c different subpopulations where
( ; )k
f x θ is the density function of the kth subpopulation with an unknown vector of
parameters, θ , and let ),...,( ,1 cPPP = be a partition of X . Then, the joint density of X is
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 757
1
( ; )i k
c
k i k
k x P
f x θ= ∈
∏ ∏ . In the CML method, ),...,( ,1 cPPP = and ),...,( 1 cθθθ = are estimated
by maximizing the log likelihood function1
ln ( ; )i k
c
k i k
k x P
f x θ= ∈
∑∑ , so the CML objective
function CMLL for the grouped data ,..., 1 nxxX = is
1
( , ; ) ln ( ; )i k
c
CML k i k
k x P
L P X f xθ θ= ∈
=∑ ∑
If a d-variate Gaussian mixture model with mean vectors, ka , and covariance matrices,
kΣ , is considered, then the CML objective function, CMLL , becomes
1
1 1
1 1( , , ; ) ln(2 ) ln ( ) ( )
2 2 2i k
c cN
CML k k i k k i k
k k x P
ndL P a X n x a x aπ −
= = ∈
′Σ = − − Σ − − Σ −∑ ∑ ∑
−−−−−= ∑∑
=
−
=
c
k
kkiki
c
k
kk axaxtrnnd
1
1
1
)')((2
1ln
2
1)2ln(
2ΣΣπ
where kk Pn = is the cardinality of kP . If the sample cross-product matrix, k
W , for the
kth cluster is ( )( )i k
k i k i k
x P
W x x x x∈
′= − −∑ , then
);ˆ,ˆ,(maxargˆ XaPLP CMLP Σ=
= ∑=
c
k k
kkP
n
Wn
1
lnminarg .
This criterion is the CML originally proposed by Scott and Symons.23
Fuzzy set theory, as proposed by Zadeh,34
has been widely used in cluster analysis
since Ruspini35
introduced the notion of fuzzy c-partitions (see also Refs. 11−13). Yang25
made the fuzzy extension of the original CML to the fuzzy CML (FCML), based on
fuzzy c-partitions. This section firstly introduces the concept of hard and fuzzy c-
partitions and then the FCML method.
The CML objective function is ∑∑= ∈
=c
k Px
kikCML
ki
xfXPL
1
);(ln);,( θθ . The parameter,
0>kα , is the proportion of the kth subpopulation in the whole population with the
constraint, ∑=
=c
k
k
1
1α . The indicator functions, 1, ,
cz z… , are used with ( ) 1
jz x = if
jx P∈ and ( ) 0
jz x = if
jx P∉ for all x in X and for all 1, ,j c= … . This is generally
known as clustering X into c clusters using 1
( , , )c
z z z= … and is called the hard c-
partition of .X Therefore, the CML objective function becomes
1 1 1
( , , ; ) ln ( ; ) ( ) ln ( ; )j
c c n
CML j j i j i j j i
j x P j i
L P X f x z x f xα θ α θ α θ= ∈ = =
= =∑∑ ∑∑ .
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 758
For the extension that allows the indicator functions, ( )j i
z x , to be functions (known
as membership functions), the values are assumed to be in the interval [0, 1], such that
1
( ) 1c
j i
j
z x=
=∑ for all i
x X∈ . In this case, the data set, X , is said to have a fuzzy c-
partition, 1
( , , )c
z z z= … , according to basic idea of fuzzy sets, as proposed by Zadeh.34
Based on this fuzzy extension, Yang25
proposed the FCML objective function as follows:
1 1 1 1
( , , ; ) ln ( ; ) lnc n c n
m m
FCML ij j i ij j
j i j i
J z X z f x r zα θ θ α= = = =
= +∑∑ ∑∑
subject to 1
1c
j
j
α=
=∑ , 0j
α > and 1
1c
ij
j
z=
=∑ for 1, ,i n= … for all [0,1].ij
z ∈ The criterion,
1m > , represents the degree of fuzziness and 0r ≥ is a scale parameter for the penalty
term 1 1
lnc n
m
ij j
j i
z α= =
∑∑ . Therefore, the FCML procedure25
is created by choosing a fuzzy c-
partition, z , and a proportion, α , and an estimate, θ , that maximize ( , , ; )FCML
J z Xα θ .
Yang25
used a multivariate normal distribution with an identity covariance matrix and
then derived a penalized fuzzy c-means clustering (PFCM) algorithm, where the PFCM
objective function is exactly the FCM objective function with the addition of the penalty
term 1 1
lnc n
m
ij j
j i
z α= =
∑∑ . There are several applications of the FCML,29-33
but the FCML
algorithm still lacks some robustness, in terms of its inability to detect different sizes and
shapes of clusters, its heavy dependence on parameter initializations and also the
necessity to provide an a priori number of clusters. In the next section, a robust
framework for FCML clustering is constructed.
3. The Proposed robust FCML Clustering Framework
Although the CML is a probability model-based cluster analysis that is used to
understand some existing clustering methods, such as agglomerative hierarchical
clustering,36
k-means7,8
and the criterion of Friedman and Rubin,37
the CML does not
allow the specification of which feature (orientation, size and shape) is common to all
clusters and which may differ between clusters. In order to address these drawbacks of
the CML, Banfield and Raftery26
first created a so-called model-based clustering by
extending the CML with the eigenvalue decomposition of the covariance matrix.
Banfield and Raftery’s26
eigenvalue decomposition approach is firstly narrated.
Since the CML originally proposed by Scott and Symons23
assumes that all the
components (clusters) are different, with a Gaussian mixture model, where
);ˆ,ˆ,(maxargˆ XaPLP CMLP Σ=
= ∑=
c
k k
kkP
n
Wn
1
lnminarg , then in order to allow the
specification of which feature is common to all clusters and which may differ between
clusters, Banfield and Raftery26
re-parameterized the covariance matrix, kΣ , in terms of
its eigenvalue decomposition with
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 759
kkkk DD ′= ΛΣ kkkk DAD ′= λ
where )(max kk diag Λλ = determines the size of the kth cluster, kD is an orthogonal
matrix representing its orientation, and 2
(1, , , )k k kd
A diag a a= … represents its shape. If
different constrains for its eigenvalue decomposition are given, the criterion for partition,
P , in the CML becomes the following clustering criteria:
1. When 2σλ =k , IAk = and IDk = for all ck , ,1…= and the clustering is spherical
and the same size as for ∑∑∑= ∈=
−==c
k Px
kiP
c
k
kP
ki
axWtrP
1
2
1
ˆminarg)(minargˆ . This
criterion was originally proposed by Ward36
.
2. When 2σλ =k , AAk = and DDk = for all ck , ,1…= and the clustering is
ellipsoidal and the same size, shape and orientation as for ∑=
=c
k
kP WP
1
minargˆ . This
criterion was originally proposed by Friedman and Rubin.37
3. When there is no constrain to kΣ for all ck , ,1…= , the clustering is ellipsoidal and
different in size, shape and orientation to that for
= ∑=
c
k k
kkP
n
WnP
1
lnminargˆ . This
criterion was the original type proposed by Scott and Symons.23
4. When 2kk σλ = , IAk = and IDk = for all ck , ,1…= and the clustering is spherical
and different in size to that for
= ∑
=
c
k k
kkP
n
WtrnP
1
lnminargˆ . This criterion was
proposed by Banfield and Raftery.26
5. When there is only the constraint, AAk = , where A is known, for all ck , ,1…= ,
each cluster is different in size and orientation but has the same known shape, so
= ∑
=
−c
k k
kkP
n
AtrnP
1
1
lnminargˆ Ω . This criterion was also proposed by Banfield
and Raftery.26
For ,..., 1 nxxX = as a random sample of size, n , from a mixture of probability
density functions, ( )θα ,;xf , and the fuzzy c-partition, 1
( , , )c
z z z= … , with values in the
interval [0, 1], such that 1
( ) 1c
j i
j
z x=
=∑ for all i
x X∈ , Yang16
proposed the FCML
objective function as follows:
1 1 1 1
( , , ; ) ln ( ; ) lnc n c n
m m
FCML ij j i ij j
j i j i
J z X z f x r zα θ θ α= = = =
= +∑∑ ∑∑
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 760
subject to 1
1c
j
j
α=
=∑ , 0j
α > and 1
1c
ij
j
z=
=∑ for 1, ,i n= … for all [0,1]ij
z ∈ . The
eigenvalue decomposition concept is used to advance the FCML model by re-
parameterizing the covariance matrix, j
Σ , in terms of its eigenvalue decomposition with
.j j j j j j j j
D D D A Dλ′ ′Σ = Λ =
The distribution, ( ; )j i
f x θ , is a multivariate normal distribution. The decomposition of
j j j j jD A Dλ ′Σ = produces the following five cases.
Case 1: If the parameters of j
Σ are allowed to vary, then
1
1/ 2/ 2
1 1 1 1
( ) ( )1( , , , ; ) ln( exp( )) ln .
2(2 )
tc n c ni j j i jm m
FCML ij ij jd
j i j ij
x u x uJ z u X z r zα α
π
−
= = = =
− Σ −Σ = − +
Σ∑∑ ∑∑ If
11( )(ln ( ) ( ) ln(2 )) ln
2
t
ij j i j j i j jd x u x u d rπ α−= − Σ + − Σ − + + , then the lagrangian FCML
J
of FCML
J is 1 2 1 2
1 1 1 1
( , , , , , ) 1 1c n c c
m
FCML ij ij j ij
j i j j
J z u z d zα λ λ α= = = =
Σ = − − − −
∑∑ ∑ ∑ . The first
derivatives of the lagrangian, FCML
J , are taken with respect to the parameters,
, , ,ij j j j
z uα Σ , and these are set to 0. Therefore, the necessary conditions for the
maximization ( , , , )z uα Σ of ( , , , )FCML
J z uα Σ can be derived as follows:
1 1 1
, 1, ,n c n
m m
j ij il
i l i
z z j cα= = =
= =∑ ∑∑ … (1)
1 1
, 1, ,n n
m m
j ij i ij
i i
u z x z j c= =
= =∑ ∑ … (2)
11
1
1
, 1, , , 1, ,c m
ij
ij
l li
dz j c i n
d
−
−
=
= = =
∑ … … (3)
1
1
( )( )
, 1, ,
nm t
ij i j i j
i
j nm
ij
i
z x u x u
j c
z
=
=
− −
Σ = =∑
∑… (4)
The factor, r , of the penalty term, 1 1
lnc n
m
ij j
j i
z α= =
∑∑ , in the FCML objective function,FCML
J ,
is now considered. Since the penalty term is a function of the mixing proportions, j
α ,
the factor, r , must be proportional to the impact of the mixing proportions, j
α , on
FCMLJ . In fact, when r = 0 , the penalty term,
1 1
lnc n
m
ij j
j i
z α= =
∑∑ , has no impact onFCML
J , but
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 761
when r = 1 , the mixing proportions, [0,1]j
α ∈ , with 1
1c
j
j
α=
=∑ , have a full impact on
FCMLJ . In this sense, a constraint for r can be set between 0 and 1, i.e. [ , ]r ∈ 0 1 . A way
to estimate the factor, r , is then proposed. Since the true values of the mixing
proportions, j
α , depend on the cluster structure of the data set, the value of r can be
estimated according to its data structure. Because [ , ]r ∈ 0 1 , the correlation coefficients
between data attributes of the data set can be used to estimate r , using the following
process. Let ,..., 1 nxxX = be a data set with n observations and d attributes. We
consider ( ,..., ), , ,s s ns
y x x s d= = …1
1 as the data vector of sth attribute for the data set,
,..., 1 nxxX = . The correlation coefficients between these attribute vectors with
( )( ) ( ) ( ),,
s t
n n n
y y is s it t is s it t
i i i
y y y y y y y y s tρ= = =
= − − − − ≠∑ ∑ ∑2 2
1 1 1
are then calculated. The initial
value of r is 1/2 and the value is increased for positive correlation coefficients and
decreased for negative correlation coefficients. Therefore, the following estimate is
proposed for r :
,1
2
s ty y
s t tr
ρ≠
+
=∑∑
(5)
Since the value of r may exceed 1, if 1r ≥ , we set 1r = . In contrast, the value of r
may become less than 0, so if 0r ≤ , we set 0r = . In this case, the factor, r , is an
aggregate of the degrees of correlation between data attributes for the given data set. If
the data set has a larger aggregated degree of correlation between attribute vectors, then
the mixing proportions, j
α , have a larger impact on the FCML objective function,
FCMLJ .
Case 2: If the covariance matrix is replaced with j
IλΣ = , then all clusters are the same
size. In this case, the objective function becomes the penalized FCM. The updated Eq. (4)
is replaced with the following equation:
( )1 1
1 1
( )( )c n
m t
ij j j j j
j i
c nm
il
l i
z tr x u x u
d z
λ= =
= =
− −
=
∑∑
∑∑ (6)
Case 3: If the cluster sizes are changed to j j
IλΣ = , the updated Eq. (4) is replaced with
the following equation:
( )1
1
( )( )n
m t
ij j j j j
i
j nm
ij
i
z tr x u x u
d z
λ =
=
− −
=∑
∑ (7)
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 762
Case 4: All clusters are the same size and shape, but have different orientations, so the
covariance matrix can be defined in the form,t
j j jD ADλΣ = . Let
jW be the sample cross-
product matrix for the kth cluster, written as ( )( )n
t
j ij i j i j
i
W z x x x x=
= − −∑1
. Note that
/j j
W n is the MLE of j
∑ . The eigenvalue decomposition of j
W becomes t
j j j jW L L= Ω ,
where , ,j j pjdiag w wΩ = …1 and mj
w is the mth eigenvalue of j
W . In deriving the
updated equations, j
L is treated as the estimation of j
D , so the updated Eq. (4) is
replaced with the following equations:
ij
c n cm
j
j i j
tr A d zλ −
= = =
= Ω
∑ ∑∑1
1 1 1
(8)
, 1, ,t
j j jL AL j cλ∑ = = … (9)
Case 5: When the parameters of the covariance matrix are the same as for case 4, but the
size to be changed, more general elliptical shapes are produced. The covariance matrix is
changed to t
j j j jD ADλΣ = . In this case, the updated Eq. (4) is replaced with the
following equations:
ij
c nm
j j
j i
tr A d zλ −
= =
= Ω
∑ ∑1
1 1
(10)
, 1, ,t
j j j jL AL j cλ∑ = = … (11)
Thus, we propose the robust FCML clustering algorithm as follows:
Robust FCML clustering algorithm
Step 1: Fix (1, )m ∈ ∞ , fix 2 c n≤ ≤ and fix any 0ε > .
Give an initial partition (0)z and let 1s = .
Estimate the factor r of the penalty term using Eq. (5).
Step 2: Compute ( )sα and ( )su with ( 1)sz − by (1) and (2).
Step 3: Compute ( )sΣ with ( 1)sz − and ( )su by (4) or (6) or (7) or (8) and (9) or (10) and
(11).
Step 4: Update to ( )sz with ( ) ( )
,s s
uα and ( )sΣ by (3).
Step 5: Compare ( )sz to
( 1)sz − in a convenient matrix norm.
If ( ) ( 1)s s
u u ε−− < , Stop.
Else 1s s= + , return to Step 2.
The use of these five cases for the robust FCML clustering algorithm is important.
The following notations are used to indicate the presentation of different cases, as
follows: vvv for case 1, ei for case 2, vi for case 3, eve for case 4 and vve for case 5.
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 763
After using the different models: vvv, ei, vi, eve and vve, for the robust FCML clustering
algorithm with different cluster numbers, the best model with the optimal number of
clusters must be selected. The Bayesian information criterion (BIC) is a very suitable tool
for model selection, so it is used for the model selection criterion.
The BIC is a model selection method that is based on the Bayesian method, using
the Bayes factor and posterior model probabilities. The main concept is that if there are
1, ,
kM M… models, with prior probabilities ( ), 1, ,
kp M k K= … , then, by the Bayes
theorem, ( | ) ( | ) ( ),k k k
p M X p X M p M∝ where ( | ) ( | , ) ( | )k k k k k k
p X M p X M p M dθ θ θ= ∫ and
( | )k k
p Mθ is the prior distribution of k
θ and ( | )k
p X M is known as the integrated
likelihood of the model,k
M . The Bayesian approach to model selection is to choose the
model that is most likely to be a posterior. To compare two models, 1
M and2
M , the
Bayes factor is defined as the ratio of the two integrated likelihoods,
12 1 2( | ) / ( | ).B p X M p X M= However, the main difficulty in using the Bayes factor is
the evaluation of the integral that defines the integrated likelihood. For regular
models, the integrated likelihood can be approximated by the BIC, using
ˆ2 log ( | ) 2 log ( | , ) log( ) ,k k k k k
p X M p X M v n BICθ≈ − = where k
v is the number of
independent parameters to be estimated ink
M . A large BIC score indicates strong
evidence for the corresponding model, so the BIC score can be used to compare models
with different covariance matrix parameterizations and different numbers of clusters for
the robust FCML clustering method. All BIC values are calculated from the cluster
number, 2, to the given maximum number of clusters for these five different models.
Using the plot of the BIC values, the best model with the optimal number of clusters is
easily chosen, so a robust FCML clustering framework is constructed, as shown in Fig. 1.
Fig. 1. Robust FCML clustering framework.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 764
4. Numerical Examples and Real Data Applications
In this section, we present several examples that use simulated data. The proposed
method is applied to these simulated data sets of different sizes, shapes and orientations.
One of the five models: vvv, ei, vi, eve and vve, is chosen as the best one with the
optimal number of clusters, using the proposed clustering framework. Yang25
proposed
the FCML, in which the penalized fuzzy c-means (PFCM) is derived using the FCML, so
that the well-known fuzzy c-means (FCM) become a special case of the PFCM.
However, the FCML with PFCM needs an a priori number of clusters to be assigned and
parameter initialization for the factor, r , of the penalty term. In 1978, Gustafson and
Kessel (GK) studied the effect of different cluster shapes, with the exclusion of the
spherical shape, by replacing the Euclidean distance, ( , )i j i jd x u x u= − , in the FCM
objective function with the Mahalanobis distance, ( ) ( )2
( , ) ,j
t
i j i j i j j i jAd x u x u x u A x u= − = − −
where j
A is a positive definite d d× matrix with a determinate, ( )det j jA ρ= , that is a
fixed constant. This is termed the GK algorithm. Some comparisons of the proposed
method with the FCM, PFCM and GK algorithms are given, using the artificial data sets.
For real data sets, the method uses iris data, diabetes data, and Wisconsin breast cancer
datasets.
Example 1. In the first example, a three-cluster data set with the same shape, but different sample sizes is considered. The data set has 300 data points; 200 data points in one cluster and 50 data points in two other clusters, as shown in Fig. 2. The proposed method is used for the data set and it is found that the BIC values select model ei as the best one, with an optimal cluster number of 3, as shown in Fig. 2. The clustering results for the five models: vvv, ei, vi, eve and vve, are also shown in Fig. 2. The FCM and PFCM are also used in this example. The FCM, PFCM and GK algorithms require the assignation of a cluster number. In fact, these algorithms were implemented using different cluster numbers, but there are always too many misclassifications for these different cluster numbers, except for the cluster number, 3, so only those clustering results for the FCM, PFCM and GK with the cluster number, 3, were used for the comparison, as shown in Fig. 3. It is seen that the results for the PFCM have less misclassifications than those for the FCM, but the proposed method is the best of them. Since the factor, r , of the penalty term is effective in the PFCM, the PFCM was also implemented for several values of r . The results are shown in Table 1.
Fig. 2. Data set and different clustering results and BIC values for Example 1.
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 765
Fig. 3. Clustering results of our method, FCM and PFCM for Example 1.
Table 1. Error rates of our method, PFCM and FCM for Example 1.
The next examples use data with variation in shape. The proposed method produces
good clustering results because the estimation of the factor, r , that is suggested is proper
and the BIC criterion is used to choose a suitable model.
Example 2. This example uses two data sets with line-shaped clusters. In the first data
set, the three line-shaped clusters are overlapped in the middle, as shown in the first
graph (original) on the left side of Fig. 4. The proposed method is used first for this data
set and it is found that the models, ei and vi, perform badly, but that the other models
produce good clustering results with an optimal cluster number of 3, as shown on the left
side of Fig. 4. Using the BIC values shown on the right side of Fig. 4, the model, eve, is
Methods Error rate
GK 4.3%
FCM ( PFCM with r = 0 ) 2%
PFCM with r = 0.2 1.9%
PFCM with r = 0.4 1.82%
PFCM with r = 0.6 1.74%
PFCM with r = 0.8 1.67%
PFCM with r = 1 1.67%
Our method 1%
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 766
best, with an optimal cluster number of 3. Figure 5 shows the clustering results for the
FCM, PFCM and GK, and the proposed method is obviously better than the others. In the
second data set, the three line-shaped clusters constitute an A-like character, as shown in
Fig. 6. For this data set, the models, ei and vi, perform badly, but the other models
produce correct cluster numbers and good clustering results. The model, vve, is selected
as the best one, with an optimal cluster number of 3. The proposed method is also better
than the FCM and PFCM, but GK performs well. The error rates and clustering results
are shown in Fig. 7.
Fig. 4. Data set and different clustering results and BIC values for the first data set.
Fig. 5. Error rates and clustering results of our method, FCM and PFCM for the first data set.
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 767
Fig. 6. Data set and different clustering results and BIC values for the second data set.
Fig.7. Error rates and clustering results of our method, FCM and PFCM for the second data set.
Example 3. In this example, a data set is generated from a mixture of three bivariate
normal distributions, using an unconstrained covariance matrix. The data set has 200 data
points in a spherically-shaped cluster and 50 data points each in two other clusters, as
shown in Fig. 8. Using the BIC values from the proposed method, as shown in Fig. 8, the
model, vvv, is obviously selected as the best one, with an optimal cluster number of 3.
For this data set, it is found that the model, vvv, performs very well, but some other
models do not. The error rates and the clustering results for the proposed method, FCM,
PFCM and GK are shown in Fig. 9. It is seen that the proposed method still performs
well for this data set.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 768
Fig. 8. Data set and different clustering results and BIC values for Example 3.
Fig. 9. Error rates and clustering results of our method, FCM, and PFCM for Example 3.
These examples show that the proposed method actually works well for these
simulated data sets, with different sizes, shapes and orientations. The next examples use
real data sets: iris data, diabetes data and Wisconsin breast cancer datasets.
Example 4. (Iris data38
) The Iris data is a typical test data set for many classification
techniques in pattern recognition. The data consists of 50 samples from each of three
species of iris flowers (iris setosa, iris virginica and iris versicoclor). The attributes of the
iris data are the length and the width of the sepal and the petal. The scatterplot for the Iris
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 769
data is shown in Fig. 10 (see Wikipedia). Figure 10 shows that there are two clusters in
each plot, the larger cluster containing two overlapped clusters. The proposed method
was implemented for the Iris data set. By using the random membership initial, (0)z , in
performing the proposed method, it is found that there are two groups of BIC values for
the Iris data set, as shown in Fig. 11. Both groups of BIC values always choose the
model, vve, as the best one. It is found that one produces an optimal cluster number of 2,
as shown on the left of Fig. 11, but the other produces an optimal cluster number of 3, as
shown on the right of Fig. 11. The proposed method was run with 200 times with the
random membership initial, (0)z , and approximately 70% of the runs produce a group of
BIC values with an optimal cluster number of 2 and that approximately 30% of the runs
produce a group of BIC values with an optimal cluster number of 3. For the case with an
optimal cluster number of 3, the best model, vve, can be more than 85% accurate in
correctly detecting the three clusters. In contrast, the second better model, eve, can be
more than 82% accurate in correctly detecting the three clusters.
Fig. 10. Iris data set.
Fig. 11. Two groups of BIC values for Iris data set.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 770
Example 5. (Diabetes dataset39
) This example uses, the diabetes data set, which has 145
observations and 3 attributes (Reaven and Miller39
). The meaning of the attributes is
described as follows: Glucose is the plasma glucose response to oral glucose, Insulin is
the plasma insulin response to oral glucose and Sspg is the degree of insulin resistance.
This data set is clinically clustered into 3 groups (chemical diabetes, overt diabetes, and
normal). Figure 12 is the 3D plot of the diabetes data. It is seen that these clusters are
overlapping and are clearly not spherical in shape. The triangular symbol in Fig. 12
denotes the cluster of normal patients. The structure of this cluster has a “fat middle”.
The circular and square symbols in Fig. 12 denote the clusters for chemical and overt
diabetes, respectively. The shapes of these two clusters resemble two wings. The varying
structure of this data means that many clustering methods are unsuitable. The data set
was used with the proposed method and the results were compared with those for the
FCM and PFCM, for a given cluster number of 3. Firstly, the results for BIC values using
the proposed method are shown in Fig. 13. It is seen that the best model is vvv, with an
optimal cluster number of 3. This is the cluster number used for the FCM and PFCM.
The error rates and clustering results for the proposed method, the FCM and the PFCM
are shown in Fig. 14. It is seen that the proposed method performs better than the FCM
and the PFCM for the Diabetes dataset.
Fig. 12. The 3D plot of diabetes data set.
Fig. 13. BIC values for diabetes data set after performed our method.
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 771
Fig. 14. Error rates and clustering results from our method, FCM and PFCM.
Example 6. (Wisconsin breast cancer dataset36,38
) This example uses the proposed
method to analyze the real Wisconsin breast cancer dataset, which is taken from the UCI
Machine Learning Repository38
(also see Wolberg and Mangasarian40
). There are 569
instances and 32 attributes in this two-cluster data set. The first attribute of this data is the
ID number and the second attribute is the diagnosis, which is treated as the true
classification (M = malignant, B = benign). Other attributes indicate the real-value
features which are computed for each cell nucleus: radius, texture, perimeter, area, etc.
The first two indicator attributes were initially ignored and the description of the results
using this data from the UCI website was used. The attributes: worst area, worst
smoothness and mean texture, were retained and the three-dimensional data was used
again, using the proposed method. The proper cluster model, vvv, has an optimal cluster
number of 2, as seen in Fig. 15. Using the cluster number, 2, the FCM and PFCM were
also implemented using this data. It is seen that the accuracy of the proposed method is
94.5%, which is better than that for the other clustering methods, as shown in Table 2.
Table 2. Average accuracies for different clustering methods.
Clustering
method
Our method PFCM with r = 1 FCM (PFCM with r = 0)
Optimal
accuracy
94.5% 92.1% 91.9%
Fig. 15. BIC values of our method for Wisconsin breast cancer dataset.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 772
Example 7. This example uses a three-component bivariate Gaussian mixture
distribution with the parameters (see Chatzis and Varvarigou41
),
1 2 3 1 2 3
1; (0,3) , (3,0) , ( 3,0) ;
3
T T Tu u uα α α= = = = = = −
1 2 3
2 0.5 1 0 2 0.5, , .
0.5 0.5 0 0.1 0.5 0.5
− Σ = Σ = Σ =
−
A data set is generated from this Gaussian mixture model with a sample size, n = 2400, as
shown in Fig. 16. The proposed method is firstly implemented using the data set and the
BIC values select the model, vvv, as the best one, with an optimal cluster number of 3, as
shown in Fig. 17. In order to determine the effect of noise on the proposed method, 600
noisy points that are drawn from a uniform distribution over the range [−20,20] are
added, as shown in Fig. 18. The proposed method is then implemented using this noisy
data set and the BIC values also select the model, vvv, as the best one, but with an
optimal cluster number of 4, as shown in Fig. 19. For this noisy data set, there is not good
clustering. That is, the proposed method is heavily affected by noisy points, for this noisy
data set. On adding 600 noisy points that are drawn from a uniform distribution over the
range [−10,10], as shown in Fig. 20, the proposed method was also implemented for this
noisy data set and the BIC values also select the model, vvv, as the best one, with an
optimal cluster number of 4, as shown in Fig. 21. However, for this noisy data set, there
is still good clustering, not affected by the noisy points too much. Overall, a comparison
with Chatzis and Varvarigou41
shows that the proposed method is not significantly
tolerant of noise or outliers; it may consider all noisy points as another cluster, as shown
in Fig. 21.
It is worthy of note that the proposed robust fuzzy CML clustering framework exhibits
clustering characteristics that are robustness to parameter initialization, cluster number
and different volumes of clusters, but not to noise and outliers. In the literature, there are
some studies of clustering characteristics with robustness to noise and outliers that use
multivariate t-distributions, such as those of Chatzis and Varvarigou,41,42
Peel and
McLachlan43
and Greselin and Ingrassia.44
Further research will extend this robust fuzzy
CML clustering framework to multivariate t-distributions.
-6 -4 -2 0 2 4
-3
-2
-1
0
1
2
3
4
5
6
Fig.16. Three-component bivariate Gaussian mixture data set.
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 773
-5 0 5-2
0
2
4
Original
-5 0 5-2
0
2
4
3
2 2
333
1
2
13
31
222
11
2
331
2
33
3 11
2
33333
2
1
2
31 1
2
1
2
1 1
2
133
1
222
11
2
3 1
2
1
2
113331
23
3
2
13
2
3 3
31
2
1
2
11
22
2
1
2
3 11
2
1
2
22
33 1
22
33
2
13
1
2
13 13
22
32
3
2
3 13 1 13
2
1 1
2
1
22
2
3 13
113 1
2
11 113 1 1 1
313
222
13
2
13
2
3
2
133
3
2
1
2 2
2
1
2
1 1
22
3 1 1
2 2
3
2
11
3 3
3
1
2
13
3 113 1
2 2
3
33
22
33 1
2
313
3
3
2
3
2
3 33
2
3
2
33
333
2
1
2
1 133
2
1
22
333
2
1111
2 2
31 1
3
2 22
2
1 133
3 1133
2
3
2
311
2
13
2233
2
3 2
1
22 2
23
1
2 2
31
2
1
233
13 1 11
2
3 11
2
1 13
2
22
3 3
2
3
23
1113 1
2
31
31
2
133
2
33
2
3
2 2
3
22
1
2
3
1
1
2
2
333 1
22
1
23
2
3
3
3 33 3
2 2
11
2
31
2
13
1 113 13
222
3 1
2
13
3 1113
1
2
11
2
3 1
2
3
33 11
2
3113 3
2
3 11
2
133
2
3 13 3
13 1
2
33
2
33
1 13
2
11
2
11
2
33
3 31 1
13
222
3 1 13
3 13
113
33
33
22
1
22
1
22
22
23
2
2
1
22
1
2
33
2
1
2
3
222
1
2
1 13
2
3
22
3133
2
1131
2
3 1
2 2
33
222
2
3 1
2
11
2
3
22
1
2
3 1
22
3
2
111
1
2 2
13 131
2
11
3 3
2
3 133
2
1
2
311
11
2
31
22
1 11
22
23
2
3 13 13
111
2
11 11333
313 1
22
13 133
2
33
2
3 3 13
2
3
2
133
2333 1 1
2
33
13
3
222
33
1
2
13 11
2
13
1
322
3
2
1
2
1
22
3 3311
22
3 31 1
22
13 13 11
22
313
2
1 1
222
13
223
1113
22
2
3
22
22
3
2
13
2
33 1
22
32
33 1
2 2
13
2
3 13
2
33 3
31
2
1
2
3
33
2 22
3 1
22
32
1 1 13
333 11
2
13
2
1
2
31
3 3
33
2
2
13 1
23
222
2
3
22
11
2
13
2
1
2
1 1
2
1
2
13
2
3
2
1 11
22
33 1
2
133
2
1
22
1
2
11
222
13 11
2
1111
22
131 1
33
31
3
2
3 13
22
22
2
322
3 1
2
2
313
2
3
2
3
22
33 1 1
2
1
2
3
22
3
2
3
2
3
22
1
2
1
2
31
2
1
2
3 3
2
33 113
2
11
222
33 3
2
11
2
13
2
33 1
222
1
22
331
2
3
2
1
22
1
2
13 3 1 1 1
222
3
222
113 11
2
1
2
222
1
2
22
1 11
2
3
2
2
2
1
22
13
2
3
22
3 1
22
3
3 11
2
11133 3
22
13
33
1
311
2
1 13
3
2
3 1
2
3
2
13
22
1 13
2
2 2
1
23
3 1
22
1
2
113 1
2
113
2
11
22
31 13
11 1
2
3
22
11
22
2
13 1
2 2
1
2
1333
2
3 113 1
2
3 1
2
133
131 133 3 113
222 2
2
133
3
2
1
2
1131
22
1 133
3
22
31 1
3
22
13 13113 133
13 1
3
2 22 2
1311 11
1 133
3
3
2
22
32
3 1
2
11
23
11
22 2
33
2
3
22
1 1
2
2
13
3
2
13
22
3
2
31
31
13
1311
2
11
2
13
23
3
2
13 113 11
22
1
2 2
13 13
22
33
2
33 3
31
3
1
2
3
223
3
2
3
2 22
11 11
2
3 13
2
3
2
3 111
22
1
2
3
222
3 13 1
2 2
33 3
2
1133 33
13
22
3
1
2
13
222 2
11
2
2
3333
2
3333 1 1
2
1
2
33
33
2
1
2
3
2 22 2
3 1
22
1
21
2
13
2
3
2
11
2
33
2
3 1 111 1
22
1
2
1
22
3
2
2
113
33
3 11 111
2
11 1
2 2
33113
2
13
22
33
2
3
22
113
22
2
21111
3 1
22
31
1
2
3
2
3
2
3
2
12
33
2
13 1
2
22
3
2
3 13
22
1
2
32 2
11
2
133 1
2
3
22 22
2
1
2
1 1 1
2
1
22
3 113 13
22
111
2
3
2
1
22
332
1
2 2
3
2
133
13 13 11
2
33
2
13 13
113 11
2
3 3 1
3
2
33
13
2
111
2
133 11
2
3
22
3
2
1
22
1
222
2
1 1
22
2
33 1 1
2
33
13
2
133
2
13
333
2
1 13
222
3
2
1
222
2
3 13
13
2
13 3
2
22 2
1
23
2
12
1
22
1
2
3
22
1113
2
1
2
1
322
3
2
33 1
33
2
3
222
2
1
2
22
1
3
111
3
2 2
3
2
2
3 1
3
33
2
333311
1
2
313
31 1
2
333
3
1113 1
2
1 13
2
3 1 1 11 113 11
2 2
3
3
1
2
1
2 2
13 11 1
22
1
2
1 13 1 133
2
3 13
3
22
11
2 2
33
2
3
2
3 3
2 22
111 13 13
22
3
2
1 1
22
3333
13
1
222
311
2
1
22
33 113 33 3
2
3
2
33
2
31
3 1
22
22
3
22
2
13
13
2
11
22
111
3
2
33
22
111
3 13 3 1
2
13
22
222
1
22
3 3
2
3 1
22
3 3
2
11
2
1
2
33
2
13
22
13
2
33
2
3
2
2
3 1
2
31
113
2313 1
2
33
1 1
22
13 1
2
11
22
3 1 113 133
313 3
2
1
2
11
2
113 1
2
13
2
1
2
113
2
2
1
2
11 1 13
11
2
11
2
3
2
3 13 13
2
113
3
1
2 2
1 1 1
2
1
2
2
133
3
3
1 11
222
1333
1 1133
3
2
3 113 1
1 13
22
133 13 1
2
1
2
3 13 131 1 1
2
33
2
3
3
2
3 1311
3
311
13
33 1 1
2
1
2
113
2
3 13 1
2
1
2
1 111
222
222
11
13
2
1
2
3
22
11
33 133
2
3
113
33
31
22
3
2
133
31 1 1
333
2
2
32
1
2
3
3 1111
3
3
2
31
2
3 1
22
3 1133
2
3 3
2
313 3 1 1313
2
22
33 1
22
3
12
13 113 13
3
1
222
113
3 3 1
2222
1 13
2
1
32
3
2
11
22
3
22
133 133 1
23
1 1
2 2
333 133
2
1
2 2
1113
2
3
2 2
2
13
2
1
2
3
2
1
3
22
22222
31
2 2
1
32
3 3 22 2
33
1
22
11
2
1
2
1
2
3 1113
2
1
22
2
3
311
23
11
1
2 2
1
22
13 1
2
13
3
2
1
2
2
1
2
1
222
33 3
3
22
1
2
13 313
3
2
3 13 1 113 1
2
3 3 1
2
3
2
13
33
1333
2
2
3 33 3 1
2
33 1
222
2
3 11
222
333 1
11131
2
11 11
3 3
2
3
VVV
-5 0 5-2
0
2
4
3
2 2
333
1
2
13
31
222
11
2
331
2
33
3 11
2
33333
2
1
2
31 1
2
1
2
1 1
2
133
1
222
11
2
3 1
2
1
2
113331
23
3
2
13
2
3 3
31
2
1
2
11
22
2
1
2
3 11
2
1
2
22
33 1
22
33
2
13
1
2
13 13
22
32
3
2
3 13 1 13
2
1 1
2
1
22
1
3 13
113 1
2
11 113 1 1 1
313
222
13
2
13
2
3
2
133
3
2
1
2 2
2
1
2
1 1
22
3 1 1
2 2
3
2
11
3 3
3
1
3
13
3 113 1
3 2
3
33
22
33 1
2
313
3
3
2
3
2
3 33
2
3
2
33
333
2
1
2
1 133
2
1
22
333
2
1111
2 2
31 1
3
2 23
2
1 133
3 1133
2
3
2
311
2
13
2333
2
3 2
1
22 2
23
1
2 2
31
2
1
233
13 1 11
2
3 11
2
1 13
2
22
3 3
2
1
33
1113 1
2
31
11
2
133
2
33
2
3
2 2
3
22
1
2
3
1
1
2
2
333 1
22
1
23
2
3
3
3 33 3
2 2
11
2
31
2
13
1 113 13
222
3 1
2
13
3 1113
1
2
11
3
3 1
2
3
13 11
2
3113 3
2
3 11
2
133
2
3 13 3
13 1
2
33
2
33
1 13
2
11
2
11
3
33
3 11 1
13
222
3 1 13
3 13
113
33
33
22
1
33
1
22
22
23
2
2
1
22
1
2
33
2
1
2
3
222
1
2
1 13
2
3
22
3133
2
1131
2
3 1
2 2
33
232
2
3 1
2
11
2
3
22
1
2
3 1
22
3
2
112
1
3 2
13 131
2
11
3 3
2
3 133
2
1
2
311
11
2
31
22
1 11
22
23
2
3 13 13
111
2
11 11333
313 1
22
13 133
2
33
2
3 3 13
3
3
2
133
2333 1 1
2
33
13
3
222
33
1
2
13 11
2
13
1
322
3
2
1
2
1
22
3 3311
22
3 31 1
22
13 13 11
22
313
2
1 1
122
13
223
1113
22
2
3
22
22
3
2
13
2
33 1
22
33
33 1
2 2
13
2
3 13
1
33 3
31
2
1
2
3
33
2 22
3 1
22
33
1 1 13
333 11
2
13
2
1
2
31
3 3
33
2
2
13 1
23
222
2
3
22
11
2
13
2
1
2
1 1
2
1
2
13
2
3
2
1 11
22
33 1
2
133
2
1
22
1
2
11
222
13 11
2
1111
22
131 1
33
31
3
3
3 13
23
22
2
322
3 1
2
2
313
2
3
2
3
22
33 1 1
2
1
2
3
22
3
2
3
2
3
23
1
2
1
2
31
2
1
2
3 3
2
33 113
2
11
223
33 3
2
11
2
13
2
33 1
223
1
12
331
2
3
2
1
22
1
2
13 3 1 1 1
222
3
222
113 11
2
1
2
222
1
2
22
1 11
3
3
2
3
2
1
22
13
2
3
22
3 1
22
1
3 11
2
11133 3
32
13
33
1
311
2
1 13
3
2
3 1
2
3
2
13
22
1 11
2
2 2
1
33
3 1
22
1
2
113 1
2
113
2
11
22
31 13
11 1
2
3
23
11
22
2
13 1
2 2
1
2
1333
2
3 113 1
2
3 1
2
133
131 133 3 113
222 2
2
133
3
2
1
2
1131
32
1 133
3
22
311
1
22
13 13113 133
13 1
3
2 22 1
1311 11
1 133
3
3
2
22
32
3 1
3
11
23
11
22 2
33
2
3
22
1 1
2
2
13
3
2
13
22
3
2
31
31
13
1311
2
11
2
13
33
3
2
13 113 11
21
1
2 2
13 13
22
33
2
33 3
31
3
1
2
3
223
3
2
3
2 22
11 11
2
3 13
2
3
2
3 111
22
1
2
3
222
3 13 1
2 2
33 3
2
1133 33
13
22
3
1
2
13
322 2
11
3
2
3333
2
3333 1 1
2
1
2
33
33
2
1
2
3
2 22 2
3 1
22
1
21
2
13
2
3
2
11
2
33
2
3 1 111 1
22
1
3
1
22
3
2
2
113
33
3 11 111
2
11 1
2 2
33113
2
13
22
33
2
3
22
113
33
2
31111
3 1
22
31
1
2
3
2
3
2
3
2
13
33
2
13 1
2
23
3
2
3 13
22
1
2
32 2
11
2
133 1
2
3
22 22
2
1
2
1 1 1
2
1
22
3 113 13
22
111
2
3
2
1
22
333
1
2 2
3
2
133
13 13 11
2
33
3
13 13
113 11
2
3 3 1
3
2
33
13
3
111
2
133 11
2
3
22
3
2
1
22
1
222
2
1 1
22
2
33 1 1
2
33
13
2
133
2
13
333
2
1 13
233
3
2
1
222
2
3 13
13
2
13 3
2
33 2
1
13
2
12
1
22
1
2
3
22
1113
2
1
2
1
322
3
2
33 1
33
2
3
222
2
1
2
22
1
3
111
3
2 2
3
2
2
3 1
3
33
2
3333 11
1
2
313
31 1
2
333
3
1113 1
2
1 13
2
3 1 1 11 113 11
2 2
3
1
1
2
1
2 2
13 11 1
22
1
2
1 13 1 133
2
3 11
3
22
11
2 2
33
2
3
2
3 3
2 22
111 13 13
22
3
2
1 1
22
3333
13
1
222
311
2
1
22
13 113 33 3
2
3
2
33
2
31
3 1
22
22
3
12
2
13
13
2
11
22
111
3
2
33
22
111
3 13 3 1
2
13
22
222
1
23
3 3
2
3 1
22
3 1
2
11
2
1
2
33
2
13
22
13
2
33
2
3
3
2
3 1
2
31
113
2313 1
2
33
1 1
22
13 1
2
11
22
3 1 113 133
313 3
2
1
2
11
2
113 1
2
13
2
1
2
113
2
2
1
2
11 1 13
11
2
11
2
3
2
3 13 13
2
113
3
1
2 2
1 1 1
2
1
2
2
133
3
3
1 11
222
1333
1 1133
3
2
3 113 1
1 13
22
133 13 1
2
1
2
3 13 131 1 1
2
33
2
3
3
2
3 1311
3
311
13
33 1 1
2
1
2
113
2
3 13 1
2
1
2
1 111
222
2 22
1113
2
1
2
3
22
11
33 133
2
3
113
33
31
23
3
2
133
31 1 1
333
2
2
32
1
2
3
3 1111
3
3
2
31
2
3 1
22
3 1133
2
3 3
2
313 3 1 1313
3
22
33 1
22
3
13
13 113 13
3
1
222
113
3 3 1
2222
1 13
2
1
32
3
2
11
22
3
22
133 133 1
23
1 1
2 2
133 133
2
1
2 2
1113
3
3
2 2
2
13
2
1
2
3
2
1
3
22
22222
31
2 2
1
32
3 3 22 2
33
1
22
11
2
1
3
1
2
3 1113
2
1
22
2
3
211
23
11
1
2 2
1
22
13 1
2
13
2
2
1
2
2
1
2
1
222
33 3
3
23
1
2
13 313
3
2
3 13 1 113 1
2
3 3 1
2
3
2
13
33
1333
2
2
3 33 3 1
2
33 1
222
3
3 11
222
333 1
11131
2
11 11
3 3
2
3
EI
-5 0 5-2
0
2
4
3
2 2
333
1
2
13
31
223
11
2
331
2
33
3 11
2
33333
2
3
2
31 1
2
1
2
1 1
2
133
1
222
11
2
3 1
2
1
2
113331
23
3
2
13
2
3 3
31
2
1
2
11
22
2
1
2
3 11
2
1
2
22
33 1
22
33
2
13
1
2
13 13
22
33
3
2
3 13 1 13
2
1 1
2
1
22
2
3 13
113 1
2
13 113 1 1 1
313
222
13
2
13
2
3
2
133
3
2
1
2 2
2
1
2
1 1
22
3 1 1
3 2
3
2
11
3 3
3
1
3
13
3 113 1
3 2
3
33
22
33 1
2
313
3
3
2
3
2
3 33
2
3
2
33
333
2
1
2
1 133
2
1
22
333
2
1111
2 2
31 1
3
2 23
2
1 133
3 1133
2
3
2
311
2
13
2333
2
3 2
1
23 2
23
1
2 2
31
2
1
233
13 1 11
2
3 11
2
1 13
2
22
3 3
2
1
33
1113 1
2
31
31
2
133
2
33
2
3
3 2
3
22
1
2
3
1
1
2
2
333 1
22
1
23
2
3
3
3 33 3
2 2
11
2
31
2
13
1 113 13
222
3 1
2
13
3 1113
1
2
11
3
3 1
2
3
33 11
2
3113 3
2
3 11
2
133
2
3 13 3
13 1
2
33
2
33
1 13
2
11
2
11
3
33
3 31 1
13
222
3 1 13
3 13
113
33
33
22
1
33
1
22
22
23
2
2
1
22
1
2
33
2
1
2
3
222
1
2
3 13
2
3
22
3133
2
1131
2
3 1
3 2
33
232
2
3 1
2
11
2
3
22
1
2
3 1
22
3
2
112
1
3 2
13 131
2
11
3 3
2
3 133
2
1
2
311
11
2
33
22
3 11
22
23
2
3 13 13
111
2
11 11333
313 1
22
13 133
2
33
2
3 3 13
3
3
2
133
2333 1 1
2
33
13
3
222
33
1
2
13 11
2
13
1
322
3
2
1
2
1
22
3 3311
22
3 31 1
22
13 13 11
22
313
2
1 1
222
13
223
1113
22
2
3
22
22
3
2
13
2
33 1
22
33
33 1
2 2
13
2
3 13
2
33 3
31
2
1
2
3
33
2 22
3 1
22
33
1 1 13
333 11
2
13
2
1
2
31
3 3
33
2
2
13 1
23
222
2
3
22
11
2
13
3
1
2
1 1
2
1
2
13
2
3
2
1 11
22
33 1
2
133
2
1
23
1
2
11
222
13 11
2
1111
22
131 1
33
31
3
3
3 13
23
22
2
322
3 1
2
3
313
2
3
2
3
22
33 1 1
2
1
2
3
22
3
2
3
2
3
23
1
2
1
2
31
2
1
2
3 3
2
33 133
2
11
233
33 3
2
11
2
13
2
33 1
223
1
22
331
2
3
2
1
22
1
2
13 3 1 1 1
222
3
222
113 11
2
1
2
222
1
2
22
1 11
3
3
2
3
2
1
22
13
2
3
22
3 1
22
3
3 11
2
11133 3
32
13
33
1
311
2
1 13
3
2
3 1
2
3
2
13
22
3 13
2
2 2
1
33
3 1
22
1
2
113 1
2
113
2
11
22
31 13
11 1
2
3
23
11
22
2
13 1
2 2
1
2
2333
2
3 113 1
2
3 1
2
133
131 133 3 113
222 2
2
133
3
2
1
2
1131
32
1 133
3
22
31 1
1
22
13 13113 133
13 1
3
3 22 2
1311 11
1 133
3
3
2
22
32
3 1
3
11
23
12
22 2
33
2
3
22
1 1
2
2
13
3
2
13
22
3
2
31
31
23
1311
2
11
2
13
33
3
2
13 113 11
22
1
2 2
13 13
22
33
2
33 3
31
3
1
2
3
223
3
2
3
2 22
11 11
2
3 13
2
3
2
3 111
22
1
2
3
222
3 13 1
2 2
33 3
2
1133 33
13
22
3
1
2
13
322 2
11
3
2
3333
2
3333 1 1
2
1
2
33
33
2
1
2
3
2 22 2
3 1
22
1
31
2
13
2
3
2
11
2
33
2
3 1 111 1
22
1
3
1
22
3
2
2
113
33
3 11 111
2
11 1
2 2
33113
2
13
22
33
2
3
22
113
33
2
31111
3 1
22
31
1
2
3
2
3
2
3
2
13
33
2
13 1
2
23
3
2
3 13
22
1
2
32 2
11
2
133 1
3
3
22 22
2
1
2
2 1 1
2
1
22
3 113 13
22
111
2
3
3
1
22
333
1
2 2
3
2
133
13 13 11
2
33
3
13 13
113 11
2
3 3 1
3
2
33
13
3
111
2
133 11
2
3
22
3
2
1
22
1
222
2
1 1
22
2
33 1 1
2
33
13
2
133
2
13
333
2
1 13
333
3
2
1
222
2
3 13
13
2
13 3
2
33 2
1
23
3
13
1
22
1
2
3
22
1113
2
1
3
1
322
3
2
33 1
33
2
3
222
2
1
2
22
1
3
111
3
2 2
3
2
2
3 1
3
33
2
333311
1
2
313
31 1
2
333
3
1113 1
2
1 13
2
3 1 1 11 113 11
2 2
3
1
1
2
1
2 2
13 11 1
22
1
2
1 13 1 133
2
3 13
3
22
11
2 2
33
2
3
2
3 3
2 22
111 13 13
22
3
2
1 1
22
3333
13
1
222
311
2
1
22
33 113 33 3
2
3
2
33
2
31
3 1
22
22
3
22
2
13
13
2
11
22
111
3
2
33
33
111
3 13 3 1
2
13
22
222
1
23
3 3
2
3 1
22
3 3
2
11
2
1
2
33
2
13
22
13
2
33
2
3
3
2
3 1
2
31
113
2313 1
2
33
1 1
22
13 1
2
11
22
3 1 113 133
313 3
2
1
2
11
2
113 1
2
13
2
1
2
113
2
2
1
2
11 1 13
11
2
11
2
3
2
3 13 13
2
113
3
1
3 2
1 1 1
2
1
2
2
133
3
3
1 11
222
1333
1 1133
3
2
3 113 1
1 13
22
133 13 1
2
1
2
3 13 131 1 1
2
33
2
3
3
2
3 1311
3
311
13
33 1 1
2
1
2
113
2
3 13 1
2
1
2
1 111
222
222
11
13
2
1
2
3
22
11
33 133
2
3
113
33
31
23
3
2
133
31 1 1
333
2
2
32
1
2
3
3 1111
3
3
2
31
2
3 1
22
3 1133
2
3 3
2
313 3 1 1313
3
22
33 1
22
3
13
13 113 13
3
1
222
113
3 3 1
2232
1 13
2
1
32
3
2
11
22
3
22
133 133 3
23
1 1
2 2
333 133
2
1
2 2
1113
3
3
2 2
2
13
2
1
2
3
2
1
3
22
22222
31
2 2
1
32
3 3 22 2
33
1
22
11
2
1
3
1
2
3 1113
2
1
23
2
3
311
23
11
1
2 2
1
22
13 1
2
13
3
2
1
2
2
1
2
1
222
33 3
3
23
1
2
13 313
3
2
3 13 1 113 1
2
3 3 1
2
3
2
13
33
1333
2
2
3 33 3 1
2
33 1
222
3
3 11
222
333 1
11131
2
11 11
3 3
2
3
VI
-5 0 5-2
0
2
4
3
2 2
333
1
2
13
31
222
11
2
331
2
33
3 11
2
33333
2
1
2
31 1
2
1
2
1 1
2
133
1
222
11
2
3 1
2
1
2
113331
23
3
2
13
2
3 3
31
2
1
2
11
22
2
1
2
3 11
2
1
2
22
33 1
22
33
2
13
1
2
13 13
22
32
3
2
3 13 1 13
2
1 1
2
1
22
2
3 13
113 1
2
11 113 1 1 1
313
222
13
2
13
2
3
2
133
3
2
1
2 2
2
1
2
1 1
22
3 1 1
2 2
3
2
11
3 3
3
1
2
13
3 113 1
2 2
3
33
22
33 1
2
313
3
3
2
3
2
3 33
2
3
2
33
333
2
1
2
1 133
2
1
22
333
2
1111
2 2
31 1
3
2 22
2
1 133
3 1133
2
3
2
311
2
13
2233
2
3 2
1
22 2
23
1
2 2
31
2
1
233
13 1 11
2
3 11
2
1 13
2
22
3 3
2
1
23
1113 1
2
31
31
2
133
2
33
2
3
2 2
3
22
1
2
3
1
1
2
2
333 1
22
1
23
2
3
3
3 33 3
2 2
11
2
31
2
13
1 113 13
222
3 1
2
13
3 1113
1
2
11
2
3 1
2
3
33 11
2
3113 3
2
3 11
2
133
2
3 13 3
13 1
2
33
2
33
1 13
2
11
2
11
2
33
3 31 1
13
222
3 1 13
3 13
113
33
33
22
1
22
1
22
22
13
2
2
1
22
1
2
33
2
1
2
3
222
1
2
1 13
2
3
22
3133
2
1131
2
3 1
2 2
33
222
2
3 1
2
11
2
3
22
1
2
3 1
22
3
2
111
1
2 2
13 131
2
11
3 3
2
3 133
2
1
2
311
11
2
31
22
1 11
22
23
2
3 13 13
111
2
11 11333
313 1
22
13 133
2
33
2
3 3 13
2
3
2
133
2333 1 1
2
33
13
3
222
33
1
2
13 11
2
13
1
322
3
2
1
2
1
22
3 3311
22
3 31 1
22
13 13 11
22
313
2
1 1
222
13
223
1113
22
2
3
22
22
3
2
13
2
33 1
22
32
33 1
2 2
13
2
3 13
2
33 3
31
2
1
2
3
33
2 22
3 1
22
32
1 1 13
333 11
2
13
2
1
2
31
3 3
33
2
2
13 1
23
222
2
3
22
11
2
13
2
1
2
1 1
2
1
2
13
2
3
2
1 11
22
33 1
2
133
2
1
22
1
2
11
222
13 11
2
1111
22
131 1
33
31
3
2
3 13
22
22
2
322
3 1
2
2
313
2
3
2
3
22
33 1 1
2
1
2
3
22
3
2
3
2
3
22
1
2
1
2
31
2
1
2
3 3
2
33 113
2
11
222
33 3
2
11
2
13
2
33 1
222
1
22
331
2
3
2
1
22
1
2
13 3 1 1 1
222
3
222
113 11
2
1
2
222
1
2
22
1 11
2
3
2
2
2
1
22
13
2
3
22
3 1
22
3
3 11
2
11133 3
22
13
33
1
311
2
1 13
3
2
3 1
2
3
2
13
22
1 13
2
2 2
1
23
3 1
22
1
2
113 1
2
113
2
11
22
31 13
11 1
2
3
22
11
22
2
13 1
2 2
1
2
1333
2
3 113 1
2
3 1
2
133
131 133 3 113
222 2
2
133
3
2
1
2
1131
22
1 133
3
22
311
3
22
13 13113 133
13 1
3
2 22 2
1311 11
1 133
3
3
2
22
32
3 1
2
11
23
11
22 2
33
2
3
22
1 1
2
2
13
3
2
13
22
3
2
31
31
13
1311
2
11
2
13
23
3
2
13 113 11
22
1
2 2
13 13
22
33
2
33 3
31
3
1
2
1
223
3
2
3
2 22
11 11
2
3 13
2
3
2
3 111
22
1
2
3
222
3 13 1
2 2
33 3
2
1133 33
13
22
3
1
2
13
222 2
11
2
2
3333
2
3333 1 1
2
1
2
33
33
2
1
2
3
2 22 2
3 1
22
1
21
2
13
2
3
2
11
2
33
2
3 1 111 1
22
1
2
1
22
3
2
2
113
33
3 11 111
2
11 1
2 2
33113
2
13
22
33
2
3
22
113
22
2
31111
3 1
22
31
1
2
3
2
3
2
3
2
12
33
2
13 1
2
22
3
2
3 13
22
1
2
32 2
11
2
133 1
2
3
22 22
2
1
2
1 1 1
2
1
22
3 113 13
22
111
2
3
2
1
22
332
1
2 2
3
2
133
13 13 11
2
33
2
13 13
113 11
2
3 3 1
3
2
33
13
2
111
2
133 11
2
3
22
3
2
1
22
1
222
2
1 1
22
2
33 1 1
2
33
13
2
133
2
13
333
2
1 13
222
3
2
1
222
2
3 13
13
2
13 3
2
22 2
1
23
2
12
1
22
1
2
3
22
1113
2
1
2
1
322
3
2
33 1
33
2
3
222
2
1
2
22
1
3
111
3
2 2
3
2
2
3 1
3
33
2
3333 11
1
2
313
31 1
2
333
3
1113 1
2
1 13
2
3 1 1 11 113 11
2 2
3
1
1
2
1
2 2
13 11 1
22
1
2
1 13 1 133
2
3 11
3
22
11
2 2
33
2
1
2
3 3
2 22
111 13 13
22
3
2
1 1
22
3333
13
1
222
311
2
1
22
13 113 33 3
2
3
2
33
2
31
3 1
22
22
3
22
2
13
13
2
11
22
111
3
2
33
22
111
3 13 3 1
2
13
22
222
1
22
3 3
2
3 1
22
3 1
2
11
2
1
2
33
2
13
22
13
2
33
2
3
2
2
3 1
2
31
113
2313 1
2
33
1 1
22
13 1
2
11
22
3 1 113 133
313 3
2
1
2
11
2
113 1
2
13
2
1
2
111
2
2
1
2
11 1 13
11
2
11
2
3
2
3 13 13
2
113
3
1
2 2
1 1 1
2
1
2
2
133
3
3
1 11
222
1333
1 1133
3
2
3 113 1
1 13
22
133 13 1
2
1
2
3 13 131 1 1
2
33
2
3
3
2
3 1311
3
311
13
33 1 1
2
1
2
113
2
3 13 1
2
1
2
1 111
222
2 22
1113
2
1
2
3
22
11
33 133
2
3
113
33
31
22
3
2
133
31 1 1
333
2
2
32
1
2
3
3 1111
3
3
2
31
2
3 1
22
3 1133
2
3 3
2
313 3 1 1313
2
22
33 1
22
3
12
13 113 13
3
1
222
113
3 3 1
2222
1 13
2
1
32
3
2
11
22
3
22
133 133 1
23
1 1
2 2
133 133
2
1
2 2
1113
2
3
2 2
2
13
2
1
2
3
2
1
3
22
22222
31
2 2
1
32
3 3 22 2
33
1
22
11
2
1
2
1
2
3 1113
2
1
22
2
3
111
23
11
1
2 2
1
22
13 1
2
13
1
2
1
2
2
1
2
1
222
33 3
3
22
1
2
13 313
3
2
3 13 1 113 1
2
3 3 1
2
3
2
13
33
1333
2
2
3 33 3 1
2
33 1
222
2
3 11
222
333 1
11131
2
11 11
3 3
2
3
EVE
-5 0 5-2
0
2
4
3
2 2
333
1
2
13
31
222
11
2
331
2
33
3 11
2
33333
2
1
2
31 1
2
1
2
1 1
2
133
1
222
11
2
3 1
2
1
2
113331
23
3
2
13
2
3 3
31
2
1
2
11
22
2
1
2
3 11
2
1
2
22
33 1
22
33
2
13
1
2
13 13
22
32
3
2
3 13 1 13
2
1 1
2
1
22
2
3 13
113 1
2
13 113 1 1 1
313
222
13
2
13
2
3
2
133
3
2
1
2 2
2
1
2
1 1
22
3 1 1
2 2
3
2
11
3 3
3
1
2
13
3 113 1
2 2
3
33
22
33 1
2
313
3
3
2
3
2
3 33
2
3
2
33
333
2
1
2
1 133
2
1
22
333
2
1111
2 2
31 1
3
2 22
2
1 133
3 1133
2
3
2
311
2
13
2233
2
3 2
1
22 2
22
1
2 2
31
2
1
233
13 1 11
2
3 11
2
1 13
2
22
3 3
2
3
23
1113 1
2
31
31
2
133
2
33
2
3
2 2
3
22
1
2
3
1
1
2
2
333 1
22
1
23
2
3
3
3 33 3
2 2
11
2
31
2
13
1 113 13
222
3 1
2
13
3 1113
1
2
11
2
3 1
2
3
33 11
2
3113 3
2
3 11
2
133
2
3 13 3
13 1
2
33
2
33
1 13
2
11
2
11
2
33
3 31 1
13
222
3 1 13
3 13
113
33
33
22
1
22
1
22
22
23
2
2
1
22
1
2
33
2
1
2
3
222
1
2
1 13
2
3
22
3133
2
1131
2
3 1
2 2
33
222
2
3 1
2
11
2
3
22
1
2
3 1
22
3
2
113
1
2 2
13 131
2
11
3 3
2
3 133
2
1
2
311
11
2
33
22
1 11
22
23
2
3 13 13
111
2
11 11333
313 1
22
13 133
2
33
2
3 3 13
2
3
2
133
2333 1 1
2
33
13
3
222
33
1
2
13 11
2
13
1
322
3
2
1
2
1
22
3 3311
22
3 31 1
22
13 13 11
22
313
2
1 1
222
13
223
1113
22
2
3
22
22
3
2
13
2
33 1
22
32
33 1
2 2
13
2
3 13
2
33 3
31
2
1
2
3
33
2 22
3 1
22
32
1 1 13
333 11
2
13
2
1
2
31
3 3
33
2
2
13 1
23
222
2
3
22
11
2
13
2
1
2
1 1
2
1
2
13
2
3
2
1 11
22
33 1
2
133
2
1
22
1
2
11
222
13 11
2
1111
22
131 1
33
31
3
2
3 13
22
22
2
322
3 1
2
2
313
2
3
2
3
22
33 1 1
2
1
2
3
22
3
2
3
2
3
22
1
2
1
2
31
2
1
2
3 3
2
33 113
2
11
222
33 3
2
11
2
13
2
33 1
222
1
22
331
2
3
2
1
22
1
2
13 3 1 1 1
222
3
222
113 11
2
1
2
222
1
2
22
1 11
2
3
2
2
2
1
22
13
2
3
22
3 1
22
3
3 11
2
11133 3
22
13
33
1
311
2
1 13
3
2
3 1
2
3
2
13
22
1 13
2
2 2
1
23
3 1
22
1
2
113 1
2
113
2
11
22
31 13
11 1
2
3
22
11
22
2
13 1
2 2
1
2
1333
2
3 113 1
2
3 1
2
133
131 133 3 113
222 2
2
133
3
2
1
2
1131
22
1 133
3
22
31 1
3
22
13 13113 133
13 1
3
2 22 2
1311 11
1 133
3
3
2
22
32
3 1
2
11
23
11
22 2
33
2
3
22
1 1
2
2
13
3
2
13
22
3
2
31
31
13
1311
2
11
2
13
23
3
2
13 113 11
22
1
2 2
13 13
22
33
2
33 3
31
3
1
2
3
223
3
2
3
2 22
11 11
2
3 13
2
3
2
3 111
22
1
2
3
222
3 13 1
2 2
33 3
2
1133 33
13
22
3
1
2
13
222 2
11
2
2
3333
2
3333 1 1
2
1
2
33
33
2
1
2
3
2 22 2
3 1
22
1
21
2
13
2
3
2
11
2
33
2
3 1 111 1
22
1
2
1
22
3
2
2
113
33
3 11 111
2
11 1
2 2
33113
2
13
22
33
2
3
22
113
22
2
31111
3 1
22
31
1
2
3
2
3
2
3
2
12
33
2
13 1
2
22
3
2
3 13
22
1
2
32 2
11
2
133 1
2
3
22 22
2
1
2
1 1 1
2
1
22
3 113 13
22
111
2
3
2
1
22
332
1
2 2
3
2
133
13 13 11
2
33
2
13 13
113 11
2
3 3 1
3
2
33
13
2
111
2
133 11
2
3
22
3
2
1
22
1
222
2
1 1
22
2
33 1 1
2
33
13
2
133
2
13
333
2
1 13
222
3
2
1
222
2
3 13
13
2
13 3
2
22 2
1
23
2
12
1
22
1
2
3
22
1113
2
1
2
1
322
3
2
33 1
33
2
3
222
2
1
2
22
1
3
111
3
2 2
3
2
2
3 1
3
33
2
333311
1
2
313
31 1
2
333
3
1113 1
2
1 13
2
3 1 1 11 113 11
2 2
3
3
1
2
1
2 2
13 11 1
22
1
2
1 13 1 133
2
3 13
3
22
11
2 2
33
2
3
2
3 3
2 22
111 13 13
22
3
2
1 1
22
3333
13
1
222
311
2
1
22
33 113 33 3
2
3
2
33
2
31
3 1
22
22
3
22
2
13
13
2
11
22
111
3
2
33
22
111
3 13 3 1
2
13
22
222
1
22
3 3
2
3 1
22
3 3
2
11
2
1
2
33
2
13
22
13
2
33
2
3
2
2
3 1
2
31
113
2313 1
2
33
1 1
22
13 1
2
11
22
3 1 113 133
313 3
2
1
2
11
2
113 1
2
13
2
1
2
113
2
2
1
2
11 1 13
11
2
11
2
3
2
3 13 13
2
113
3
1
2 2
1 1 1
2
1
2
2
133
3
3
1 11
222
1333
1 1133
3
2
3 113 1
1 13
22
133 13 1
2
1
2
3 13 131 1 1
2
33
2
3
3
2
3 1311
3
311
13
33 1 1
2
1
2
113
2
3 13 1
2
1
2
1 111
222
222
11
13
2
1
2
3
22
11
33 133
2
3
113
33
31
22
3
2
133
31 1 1
333
2
2
32
1
2
3
3 1111
3
3
2
31
2
3 1
22
3 1133
2
3 3
2
313 3 1 1313
2
22
33 1
22
3
12
13 113 13
3
1
222
113
3 3 1
2222
1 13
2
1
32
3
2
11
22
3
22
133 133 1
23
1 1
2 2
333 133
2
1
2 2
1113
2
3
2 2
2
13
2
1
2
3
2
1
3
22
22222
31
2 2
1
32
3 3 22 2
33
1
22
11
2
1
2
1
2
3 1113
2
1
22
2
3
311
23
11
1
2 2
1
22
13 1
2
13
3
2
1
2
2
1
2
1
222
33 3
3
22
1
2
13 313
3
2
3 13 1 113 1
2
3 3 1
2
3
2
13
33
1333
2
2
3 33 3 1
2
33 1
222
2
3 11
222
333 1
11131
2
11 11
3 3
2
3
VVE
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7-2.15
-2.1
-2.05
-2
-1.95
-1.9
-1.85
-1.8
-1.75x 10
4
cluster numbersB
IC v
alu
es
VVV
EI
VI
EVE
VVE
Fig. 17. Data set without noises and different clustering results and BIC values for Example 7.
-25 -20 -15 -10 -5 0 5 10 15 20 25
-15
-10
-5
0
5
10
15
Fig. 18. Gaussian mixture data set with 600 noisy points from a uniform distribution on [−20,20].
-40 -20 0 20 40
-10
0
10
Original
-40 -20 0 20 40
-10
0
10
222222 22 222 22222222222 22222222 2222 22 22222222222
222 222 2222 222
22 2222
22 2222 22222 22 22 22
222 222
22 2222 222222222222222 222222
22 222222 2
222222 222222
22222 2222
2222222222
2 222222 2222 222222 22
22 222 22222 22 22222222222 222 2
222 2222222 2
2222 22222222
22 222222
2222
2222222222 22222 222 22222 222
2 22 2222 22222 222222222 2222 2
222
2 22222
2 22222
2222222 222222 2222222 22222 222 222 222 2222
22
22222 222222
222222222222222 2
22222
222 22 2222 2222222222222 2222 2222222 222 2222222222222
22 22222222 22222222222
222 2222222 22222222 222222 2222 22222 2222222222
2222 22
222 2222222 2
222222
2222 222222222222222 2222222222222
2222 222 22
2 22222
222 222 222 222 22 22 22222
22222222222 2222222222 22222222222222222222222222222 222222222
22 22222 22222 22
22 2222 2
222 2222 222222 222 22 222 22222 222 2
22222 222222 2222
2222 2222222
222222 222222
2222 2222 222222 2
22 222222
222 2222 22 222222222 222222
222222 22222222222
222222222222222 2222 22 2222 2222222 2222222222 222
222 22
222 22 22 22222 2222 222222 22222222222 22222222222222 222 22 222 2222 2
22222222222 22 2
22 222222222 22 222222222
2222 22222222
2 2222 2222 22 22 222222 2222 22
222222222
22222222 2222
222222222 222222222 22
22222222 22
222
22 222 222222 22 2
2222 22 222222 22 22222222 222 2222 2222222222 22
2 22
2 222 22222 222 22 2222 22222222 222 22222222222222
22 22222 222 2222222222
2222222 222222
222222 222 22222 2222222
2 22222222222222
24222222 22222 2222 22222
22 2222222 22 2222
222 222222
222 222 22222
222222 22 222222222222 222222222222 222
222222
2 22222222 22
22 2 22222 222222222
22 2222222222222 222 2
222
22222222 2222
2222222 22222222222 222 2
2222 2222
222222 222222222 22222 2222
22
222222222
22222 222 222 22222 22222
22222222 2222
2222222
22222 2 22
2 222222222222222222 222 2222222 22 2
2222
2 222 222222222 222222 2222222222 222 222
2222 222222222 222 2
222222 222222 2222222 2222 22222
2222222 22 22222 22222 22222222222
222
2222222
222 2222222 222222
22
222 22222222222 222 2222 22
222222222222
22 222 222222222
2222222 2222 2222222 2
22222 22
222222222222222 2
2 2222222222222
2222 22 22 222222222
2222222222222 2222
2222
2222222 222 222222 22222 22
22222 2222 22 22222 2222222 22 2222 222 222 222222222222 22222 222222 222222222
2222
2222
22 222
222222222222 222 22222222 222222222222222
222 22
2 22222222
222222222 22222 22
22 222 22 222
22222222 222 22
222222222 222 222222 222
22 22 2222 22242222 2
2222222 22
222222222 22222 22
22 22222 222222 222222 222222 222
22 22
222222
22 22222222222222 2 2222222222222 222 2222
222222222222222222
22222222
22222
22 2222 2
22222222
222
2222222 2222 222222222222 2222
222222 2
2222222222222 22 22
22222222222
222
2222222 2222 222
2 2222222 2222
222 222 22222222 222222 2222222222
222 222 222222222 22222222 2222 222
222222 22 2222 22222222 222 22
22222222 22222222 2222222 222222222 2
22222
22 22222
222222222
4
4
11
23
2
44
31
1
3
11 1
4
2
44
4
1
43
31
43
23
3 3
4
4
1
4
1
3
3
3 4
11
3
3
33
4
4
1
4
1
4
2
4
3
4
32
4
3
11
3
3
44
3
44
4 4
3
3 44
1
4
1
2
3 1
3
32
2
1
2
1 1
3
11
4
11
4
3
1
34
3
433
4
3
3
4
1
3 2
3
3
4
4
11
4 4
3
4
1
3
1
4
3
1
4
3
3
3 3
1
43
1
1
3
1
3
11
3 1
3
4
4 422
2
113
2
4
31
233
1
24
3
44
434
3
41
43
3
4
13 1
4
4
1
3
11
44
1
2 4
4 44
1
1
3
3233
3 13
3
3
3 1
3
4 433
31 1
4
1
1
4
111
1
31
1
13
12
4
3
3
21
1
4
3 1
33
13
21
33
3
1
23
3
21
4
3 2
3
13
4
23
4
42
3
1
4
1
3
2
4222
3
24
2
4
4
1
4
3 1
4
4
34
3
1
43
13
44
133
3
4
31
4
2
4
23
4
1
44
1
43
1
34
134
1
4
3
3
3
4
14
4
24
3
4
333 4
1
3
4
4
4
1
4
1
4
1
4
31
3
3
2
2
43 4
3
4
1
344
1
2 44
1
2
43
33
4
4
3 3
4 4
3 1
33
4
1
1
4
3
3
44
3 4
3 1
3
1
4
2
3
2
1
4 4
11
1
44
42
11
44
3
4
1
4
1
4
4
1
3 42
2
4
1
4
1
3 43
3
22
1
4
11
2
11
4
1
4
4
44
32
1
4
1
4
3
3
3
4
11
4
1
12
3
31
11 123
3
2
1
4
1
2
4 4
1
4
3 331
3
2113
4
12
11
3 4
1
2
4
3
4
4
1
442
4
3
4
13
4
2
334
43
13
3
13
1 13
4
133
1
3
4443
3
3
1
3 13
3
3
2
4
1
3
44
1
3
1
3
4
4
33
1
4
311
4442
3
4
4
3
3
33
1
4
1
31
43
1
4
1
1
4 4
1
413
VVV
-40 -20 0 20 40
-10
0
10
22222222 222 222222 22222 22222222 2222 2222222222222
222 222 2222 222
22 2222
22 2222 22222 22 22 22222 222
22 2222 222222222222222 222222
22 222222 2
222222 222222
22
222 22222222222222
2 222222 2222 222222 22
22 222 22222 22 22222222222 222 2
222 2222222 2
2222 22222222
22 222222
2222
2222222222 22222 222 22222 222
2 22 2222 22222 222222222 2222 2
222
2 22222
2 22222
2222222 222222 2222222 22222 222 222 2222222
22
22222 222222
222222222222222 2
22222
22222 2222 2222222222222 2222 2222222 222 2222222222222
22 22222222 22222222222
222 2222222 22222222 222222 2222 22222 22222222222222 22
222 2222222 2222222
2222 222222
222222222 22222222222222222 222 22
2 22222
222 222 222 222 22 22 22222
22222222222 2222222222 22222222222222222222222222222 222222222
22 2222222222 22
22 2222 2
222 2222 222222 222 22 222 22222 222 2
22222 222222 2222
2222 22222
22
222222 222222
2222 2222 222222 2
22 222222
222 2222 22 222222222 2222 22
222222 2222222222222222222222
2222 222222 2222 2222222 2222222222 222
222 22
222 22 22 22222 2222222222 22222222222 22222222222222 222 22 222 2222 22
2222222222 22 222 22
2222222 22 222222222
2222 22222222
2 2222 2222 22 22 222222 2222 22
222222222
22222 222 222
22
22222222 22222222222
22222222 22
222
22 222 22222222 2
2222 22 222222 22 22222222 222 2222 2222222222 22
2 22
2 222 22222 222 22 2222 22222222 222 22 222222222222
22 22222 222 2222222222
2 222222 222222
2222 22222 22 222 2
222222
2 22222 222222222
22222222 22222 2222 22222
22 2222222 22 2222
22
2 2222222
22222 22222
222222 22 222222222222 2222
22222222 222
222222
2 22 222222 22
22 2 22222 222222222
22 2222222222222 222 2
222
22222222 2222
2222222 22222222222 222 22222 2222
222222 222222222 22222 2222
22
222222222
22222 222 222 22222 22222
22222222 2222
222222222222 2 22
222222222222222 2222 222 22222 22 22 2
2222
2 222 222222222 222222 2222222222 222 222
2222 222222222 222 2
222222 222222 2222222 2222 22222
2222222 22 22222 22222 22222222222
222222222
2222 22222
22 222222
22
2 22 22222222222 222 2222 22
222222222222
22 222 222222222
2222222 2222 2222222 2
22222 22
222222222222222 2
2 22222222
222222
222 22 22 22 22222 2
22222222222222 22
222222
2222222 222 222222 22222 22
22222 2222 22 22222 2222222 22 2222 222 222 222222222222 22222 222222 222222222
2222
2222
22 222
222222222222 222 22222222222222222222222222 22
2 22222222
222222222 22222 22
22 222 22 222
22222222 222 22
222222222 222 222222 222
22 22 2222 22222222 2
2222222 22
222222222 22222 22
22 22222 222222 222222 222222 222
22 22
222222
22 22222222222222 2 222222222 2222 222 2222
22222222222 2222222
22222222
22222
22 2222 2
22 222222
222
2222222 2222 222222222222 222 2
2222 22 22
222222222222 22 222
2222222222222
2222222 2222 222
2 2222222 2222
222 222 22222222 222222 2222222222
222 222 222222222 22222222 2222 222
222222 222222 22222222 222 22
22222222 22222222 2222222 222222222 2
22222
22 22222
222222222
4
4
13
25
2
55
31
1
5
11 1
5
2
45
2
1
25
31
45
23
33
5
4
2
4
1
2
3
5 4
11
5
3
33
5
4
1
5
1
5
2
5
3
5
32
4
3
11
3
3
45
3
52
5 5
3
5 45
1
4
1
2
3 1
3
32
2
1
2
1 1
5
11
4
11
5
3
1
54
3
555
5
3
3
4
1
2 2
3
5
4
4
11
2 4
3
5
1
5
1
4
3
1
4
3
3
5 5
1
42
2
1
2
1
3
21
3 1
3
4
2 422
2
113
2
4
31
223
2
24
3
44
454
3
21
23
3
4
13 1
2
5
1
2
11
45
1
2 4
5 55
2
1
3
5233
3 13
3
3
3 1
3
5 455
31 1
4
1
1
5
111
1
32
1
33
12
4
3
5
21
1
5
3 1
55
13
21
55
3
3
23
5
21
5
3 2
3
22
2
23
5
42
5
1
5
1
5
2
2222
5
24
2
4
4
2
4
3 1
4
4
54
2
1
45
13
25
133
5
5
31
5
2
4
23
4
1
45
1
25
4
55
134
2
5
3
3
3
5
24
5
22
3
2
555 4
1
3
2
5
4
1
2
1
5
1
5
31
2
3
2
2
55 4
5
4
1
544
1
2 44
1
2
55
33
4
4
3 2
5 4
3 1
33
4
1
1
4
3
5
45
5 4
3 1
5
1
4
2
3
2
1
4 4
11
1
22
22
22
55
3
5
3
4
3
4
2
1
5 42
2
2
1
4
1
5 25
3
22
1
4
12
2
11
5
1
4
4
45
32
2
5
2
4
3
5
3
5
11
4
1
12
3
31
23 122
5
2
1
5
1
2
4 4
3
5
3 332
5
21 13
5
12
11
5 4
1
2
4
3
4
4
1
442
5
3
5
13
4
2
334
45
13
5
33
2 13
4
133
1
3
4445
3
2
1
3 15
3
3
2
2
1
3
44
1
5
4
3
4
5
35
1
2
311
4522
3
4
4
3
5
33
1
5
1
31
42
1
5
4
1
5 4
2
443
EI
-40 -20 0 20 40
-10
0
10
222222 22 222 22222222222 22222222 2222 22 22222222222
222 222 2222 222
22 2222
22 2222 22222 22 22 22
222 222
22 2222 222222222222222 222222
22 222222 2
222222 222222
22222 2222
2222222222
2 222222 2222 222222 22
22 222 22232 22 22222222222 222 2
222 2222222 2
2222 22222222
22 222222
2222
2222222222 22222 222 22222 222
2 22 2222 22222 222222222 2222 2
222
2 22222
2 22222
2222222 222222 2222222 22222 222 222 222 2222
22
22222 222222
222222222222222 2
22222
222 22 2222 2222222222222 2222 2222222 222 2222222222222
22 22222222 22222222222
222 2222222 22222222 222222 2222 22222 2222222222
2222 22
222 2222222 2
222222
2222 222222222222222 2222222222222
2222 222 22
2 22222
222 222 222 222 22 22 22222
22222222222 2222222222 22222222222222222222222222222 222222222
22 22222 22222 22
22 2222 2
222 2222 222222 222 22 222 22222 222 2
22222 222222 2222
2222 2222222
222222 222222
2222 2222 222222 2
22 222222
222 2222 22 222222223 222222
222222 22222222222
222222222222222 2222 22 2222 2222222 2222222222 222
222 22
222 22 22 22222 2222 222222 22222222222 22222222222222 222 22 222 2222 2
22222222222 22 2
22 222222222 22 222222222
2222 22222222
2 2222 2222 22 22 222222 2222 22
222222222
22222222 2222
222222222 222222222 22
22222222 22
222
22 222 222222 22 2
2222 22 222222 22 22222222 222 2222 2222222222 22
2 22
2 222 22222 222 22 2222 22222222 222 22222222222222
22 22222 222 2222222222
2222222 222222
222222 222 22222 2222222
2 22222222222222
22222222 22222 2222 22222
22 2222222 22 2222
222 222222
222 222 22222
222222 22 222222222222 222222222222 222
222222
2 22222222 22
22 2 22222 222222222
22 2222222222222 222 2
222
22222222 2222
2222222 22222222222 222 2
2222 2222
222222 222222222 22222 2222
22
222222222
22222 222 222 22222 22222
22222222 2222
2222222
22222 2 22
2 222222222222222222 222 2222222 22 2
2222
2 222 222222222 222222 2222222222 222 222
2222 222222222 222 2
222222 222222 2222222 2222 22222
2222222 22 22222 22222 22222222222
222
2222222
222 2222222 222222
22
222 22222222222 222 2222 22
222222222222
22 222 222222222
2222222 2222 2222222 2
22222 22
222222222222222 2
2 2222222222222
2222 22 22 222222222
2222222222222 2222
2222
2222222 222 222222 22222 22
22222 2222 22 22222 2222222 22 2222 223 222 222222222222 22222 222222 222222222
2222
2222
22 222
222222222222 222 22222222 222222222222222
222 22
2 22222222
222222222 22222 22
22 222 22 222
22222222 222 22
222222222 223 222222 222
22 22 2222 22242222 2
2222222 22
222222222 22222 22
22 22222 222222 222222 222222 222
22 22
222222
22 22222222222222 2 2222222222222 222 2222
222222222222222222
22222222
22222
22 2222 2
22222222
222
2222222 2222 222222222222 2222
222222 2
2222222222222 22 22
22222222222
222
2222222 2222 222
2 2222222 2222
222 222 22222222 222222 2222222222
222 222 222222222 22222222 2222 222
222222 22 2222 22222222 222 22
22222222 22222222 2222222 222222222 2
22222
22 22222
222222222
4
4
13
24
4
44
31
1
4
11 1
4
3
44
4
3
43
31
44
13
3 3
4
4
1
4
1
3
3
4 4
11
3
3
33
4
4
1
4
1
4
2
4
3
4
32
4
3
11
3
3
44
3
44
4 4
3
4 44
1
4
1
2
3 1
3
32
2
3
2
1 1
4
11
4
11
4
3
1
34
3
443
4
3
3
4
1
3 2
3
3
4
4
11
4 4
3
4
1
4
1
4
3
1
4
3
3
4 4
1
43
3
1
3
1
3
11
3 1
3
4
4 422
2
113
2
4
31
233
3
24
3
44
441
3
41
43
3
4
13 1
4
4
1
3
11
44
1
2 4
4 44
3
1
3
4233
3 13
3
3
3 1
3
4 444
31 1
4
1
1
4
111
1
31
1
33
11
4
3
4
21
1
4
3 1
44
13
21
43
3
3
23
4
21
4
3 2
3
13
4
23
4
42
4
1
4
1
4
3
4223
4
24
2
4
4
2
4
3 1
4
1
44
3
1
44
13
44
133
4
4
31
4
4
4
23
4
1
44
1
44
1
44
131
1
4
3
3
3
4
24
4
34
3
4
444 4
1
3
4
4
4
1
4
1
4
1
4
31
4
3
2
2
44 4
3
4
1
444
1
2 44
1
2
44
33
4
4
3 3
4 4
3 1
33
4
1
1
4
3
4
44
4 4
3 1
4
1
4
2
3
2
1
4 4
11
1
44
42
31
44
3
4
3
4
3
4
4
1
4 42
3
4
1
4
1
4 44
3
23
1
4
13
4
11
4
1
4
4
44
34
3
4
3
1
3
4
3
4
11
4
1
12
3
31
33 123
4
2
1
4
1
2
4 4
3
4
3 333
4
2113
4
12
11
4 4
1
2
4
3
4
4
1
442
4
3
4
13
4
2
331
43
13
4
33
1 13
4
133
3
3
4444
3
3
1
3 13
3
3
2
4
1
3
44
1
4
1
3
4
4
33
1
4
311
4442
3
4
4
3
4
33
1
4
1
31
43
1
4
1
1
4 4
3
113
VI
-40 -20 0 20 40
-10
0
10
25252525 255 222555 25225 22522252 2555 5525552255255
225 255 2555 255
52 2225
55 2225 22525 25 55 22525 555
55 2555 225222222255525 222525
55 255555 5
222255 255555
52
555 22552255555255
5 225525 2225 255525 22
55 225 22555 25 22255555525 255 2
555 5222525 2
5225 22225522
55 252252
5552
5522525555 22555 222 22255 255
5 22 2525 25552 252552525 5555 5
255
5 22225
2 22555
2222225 222552 2255255 22525 225 255 2555522
52
52255 225522
555555555522525 2
55225
52555 2225 2255552522525 2255 2255225 225 2522252225555
55 22225225 25552522255
555 2555555 22555525 222225 2555 22525 22552255525555 25
555 5555525 2522552
5525 255525
552222555 25222552252252522 225 25
2 22255
255 225 225 555 22 25 25522
55525255255 2555555555 22555255555552255552522252555 255555552
55 2525222225 25
55 2225 2
555 5555 225555 255 22 255 25555 555 2
55555 225555 2555
2555 22525
52
522525 252255
5555 2555 255525 2
55 255552
525 2255 25 255555255 2225 25
255555 5522522222255552525552
5555 222555 2555 5225525 2255222525 222
525 22
525 55 25 25225 5555255555 25555525225 25222522555552 225 25 555 2525 52
5222225555 25 555 25
5255222 25 222255225
5555 22555555
5 2225 2255 25 25 255525 2225 22
555525555
55525 255 255
25
55222525 22555255225
25555555 25
525
52 255 25522525 2
5255 25 225555 25 25555525 255 2555 2222555225 25
5 25
5 252 22525 225 25 2555 22255555 255 25 225555225525
55 22555 225 2555255555
5 525255 222525
5525 25255 55 255 2
555555
2 25252 255552555
55225255 22255 2225 55555
25 2555552 25 2225
25
5 2255555
55225 22225
522255 25 225555525555 2522
55552225 225
225555
5 25 522225 25
25 2 22225 255522522
52 2522555552255 225 2
522
55555555 2255
5555525 25525522225 222 25555 2525
252255 255552255 22555 2255
52
525552555
55255 225 255 22525 25555
52552225 2522
522222255225 5 25
225552255555552 2255 255 55525 55 25 2
5255
5 555 222252552 255252 2552255555 225 225
5555 252552555 255 2
555225 225255 2555255 2555 25555
5555555 25 25555 52555 25552252255
525252222
5225 22222
55 225255
25
5 25 55522222255 225 5255 25
552555555522
55 525 225555222
5552555 2555 5555255 2
52525 55
252255552552255 5
2 22555555
522225
255 25 55 25 25555 2
52555555222255 25
255525
5525255 255 222555 25225 22
55255 2525 25 52555 2522255 25 2225 255 255 555525255225 22255 252555 252222525
2555
5255
55 225
255555555225 255 25525222255555252255555225 25
5 25225552
555555555 25555 25
55 225 25 255
25255555 225 25
555522255 255 555525 255
25 25 2225 25255555 2
5255555 25
555225525 22255 25
55 22255 222555 225555 222555 225
55 25
255222
55 22222522252525 5 255555552 2555 525 2525
25555552555 2555555
25225252
55225
25 2525 2
55 252525
552
5255555 2225 555522522255 252 2
5555 25 25
225522225555 25 525
2522522525252
5525255 5555 252
5 2225555 5225
555 222 22525225 222555 2252555522
555 255 222555255 22555225 2225 222
552555 255225 22555555 255 22
55555525 25255555 2225555 222222522 2
52555
25 25552
555252525
4
4
11
57
2
77
31
4
7
11 1
7
5
47
2
1
56
31
67
23
35
7
4
1
5
1
5
3
7 4
11
6
3
33
6
4
1
7
1
7
5
7
3
6
32
5
3
41
3
3
47
3
65
7 7
3
7 56
4
4
1
5
3 1
6
55
5
1
5
1 1
7
11
4
11
7
3
4
64
3
776
7
3
3
5
1
5 2
3
3
4
5
11
5 5
3
6
1
7
1
2
3
4
4
3
3
7 6
1
45
5
1
5
1
3
11
3 4
3
4
6 452
5
113
5
4
31
253
5
54
5
44
264
3
24
22
3
4
43 4
2
7
1
2
11
67
1
2 2
7 66
1
1
3
6253
3 15
3
3
3 1
5
6 477
31 1
5
1
4
7
111
1
52
4
13
42
4
6
7
22
1
7
3 1
67
15
52
76
5
1
53
7
54
7
3 2
3
25
6
23
6
22
7
4
6
1
7
5
6525
7
54
5
4
4
2
5
5 1
4
4
76
6
1
57
13
56
453
7
7
34
7
2
4
53
4
1
46
1
67
4
77
454
2
7
1
3
3
7
22
6
55
3
5
777 4
1
3
6
7
2
1
5
4
7
1
7
51
6
3
5
5
77 4
3
4
1
644
1
5 45
1
2
66
33
4
4
6 5
7 4
3 1
55
4
1
4
5
3
6
46
7 6
5 4
7
1
6
5
3
5
1
6 4
11
1
26
25
12
77
3
6
1
4
1
4
6
1
6 52
5
6
1
4
1
7 57
1
55
1
2
11
2
11
6
1
4
2
47
32
1
7
1
4
5
7
3
6
11
4
1
45
3
34
11 125
7
2
1
7
1
2
5 4
1
7
3 335
7
52 43
6
22
11
7 4
1
2
4
3
2
4
1
445
6
3
6
13
4
2
354
66
13
6
13
1 45
4
133
1
5
4447
5
2
1
2 46
3
3
5
6
4
3
44
4
7
4
3
4
6
33
1
6
511
4665
3
4
4
3
7
13
4
7
1
31
46
1
7
4
1
6 4
1
443
EVE
-40 -20 0 20 40
-10
0
10
255555 25 255 22255525225 22552255 2555 55 55555255255
255 255 2555 255
55 2225
55 2225 25555 25 55 22
525 555
55 2555 225255225555525 222525
55 255555 5
222255 255555
52555 2255
5555555255
5 225525 2525 255525 22
55 225 22535 25 22255555525 255 2
555 5222555 2
5255 22225522
55 255555
5552
5525555555 52555 225 22255 255
5 15 2525 25555 252555555 5555 5
555
5 22555
5 22555
2222225 222555 2555255 22555 225 255 255 5522
52
52255 255522
555555555522525 2
55555
525 55 2225 2255552522525 2555 2255225 225 2522555225555
55 22225225 25552522255
555 2555555 25555525 222225 2555 22525 2255225552
5555 25
555 5555525 2
522552
5555 255555552222555 2522255255225
5555 225 25
5 22255
255 225 225 555 25 25 25522
55525255555 2555555555 22555255555552255552522252555 255555555
55 25555 22225 25
55 2225 2
555 5555 225555 255 25 255 25555 555 2
55555 225555 2555
2555 2252552
522525 252255
5555 2555 255525 2
55 255552
525 2255 25 255555253 222525
255555 55225522222
555525255555555 2225 55 2555 5255525 2555222525 222
525 22
525 55 25 25255 5555 255555 25555525255 25522522555552 225 25 555 2525 5
25222255555 25 5
55 255255225 25 222255225
5555 25555555
5 5225 2555 25 25 255525 2555 25
555525555
55525255 2552
555222525 225555555 25
25555555 25
525
55 255 255225 25 2
5255 25 225555 25 25555525 255 2555 2222555555 25
5 25
5 255 22525 225 25 2555 22255555 255 25255555555555
55 22555 225 2555555555
5525255 222555
552525 555 55255 2555555
5 25555255555555
54225255 25255 2225 55555
55 2555555 25 2225
255 225555
555 525 22225
552255 25 225555555555 255255552225 225
555555
5 25522525 25
25 5 22225 255552522
55 2522555552255 225 2
522
55555555 2255
5555525 25525522225 225 2
5555 2525
252555 255552555 22555 2255
52
525552555
55555 225 255 22525 25555
52551255 2552
5222222
55225 5 25
5 255522555555552255 255 5552555 25 2
5255
5 555 222252555 255255 2552255555 255 255
5555 252552555 255 2
555555 225255 2555255 2555 25555
5555555 25 25555 52555 25552252555
525
2522225
225 2252255 225555
55
555 55555555555 525 5255 55
552555555522
55 555 225555222
5552555 2555 5555255 2
52555 55
252255555552255 5
5 2255555552222
5555 25 55 252555525
5555555225255 2525
5525
5525255 255 222555 25555 22
55255 2555 25 55555 2522555 25 2225 253 555 555525255555 22255 252555 252255555
5555
5555
55 255
255555555555 255 25525225 255555552555555
255 25
5 25255552
555555555 25555 25
55 225 25 255
25555555 225 25
555522255 253 555555 255
25 25 2225 25545555 2
5255555 25
555525555 22255 25
55 22255 222555 255555 222555 225
55 25
255222
55 22222522255555 5 2555555552555 525 2555
555555525552555555
55225252
55555
25 2525 2
55252525
555
5555555 2255 555522522255 2552
555525 2
5255555225555 25 55
52522522555
552
5525255 5555 252
5 2225555 5225
555 225 22525225 222555 2252555525
555 255 222555255 25555225 2225 222
555555 25 5255 22555555 255 22
55555525 25255555 2225555 222222525 2
52555
25 25552
555255555
4
4
11
33
4
44
31
1
4
11 1
4
3
44
4
1
43
31
43
43
3 3
4
4
1
4
1
3
3
3 4
11
3
3
33
4
4
1
4
1
4
5
4
3
4
35
4
3
11
3
3
44
3
44
4 4
3
3 44
1
4
1
4
3 1
3
35
5
1
3
1 1
3
11
4
11
4
3
1
34
3
433
4
3
3
4
1
5 2
3
3
4
4
11
4 4
3
4
1
3
1
4
3
1
4
3
3
3 3
1
43
1
1
5
1
3
11
3 1
3
4
4 451
5
113
5
4
31
133
1
54
3
44
434
3
41
43
3
4
13 1
4
4
1
3
11
44
1
5 4
4 44
1
1
3
3433
3 13
3
3
3 1
3
4 433
31 1
4
1
1
4
111
1
34
1
13
14
4
3
3
51
1
4
3 1
33
13
54
33
3
1
53
4
51
4
3 5
3
15
4
13
4
42
4
1
3
1
4
5
4555
4
54
5
4
4
1
4
3 1
4
4
34
3
1
43
13
44
133
3
4
31
4
4
4
53
4
1
44
1
43
4
34
134
1
4
3
3
3
4
14
4
34
3
4
333 4
1
3
4
4
4
1
4
1
4
1
4
31
3
3
5
5
44 4
3
4
1
344
1
3 44
1
1
43
33
4
4
3 3
4 4
3 1
33
4
1
1
4
3
3
44
4 4
3 1
4
1
4
5
3
5
1
4 4
11
1
44
43
11
44
3
3
1
4
1
4
4
1
3 45
5
4
1
4
1
3 43
3
53
1
4
11
4
11
4
1
4
4
44
34
1
4
1
4
3
3
3
4
11
4
1
15
1
31
11 155
4
2
1
4
1
5
4 4
1
4
3 331
3
5113
4
15
11
4 4
1
2
4
3
4
4
1
445
4
3
4
13
4
5
334
43
13
3
13
1 13
4
133
1
3
4443
3
3
1
3 13
3
3
5
4
1
3
44
1
4
4
3
4
4
33
1
4
311
4445
3
4
4
3
3
33
1
4
1
31
43
1
4
4
1
4 4
1
443
VVE
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7-4
-3.9
-3.8
-3.7
-3.6
-3.5
-3.4
-3.3
-3.2
-3.1
-3x 10
4
cluster numbers
BIC
valu
es
VVV
EI
VI
EVE
VVE
Fig. 19. Data set with 600 noisy points and different clustering results and BIC values for Example 7.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 774
-10 -5 0 5 10
-8
-6
-4
-2
0
2
4
6
8
Fig. 20. Gaussian mixture data set with 600 noisy points from a uniform distribution on [−10,10].
-20 0 20
-5
0
5
Original
-20 -10 0 10 20
-5
0
5
2 34
24
2 324
2444
2 32 332
44333 3
43
433
4
2 34
244
34
133
43
4 4333
43
433
43
444322
4
22 332
2 2 3
4
3
4444
2444
24
244
332 324
3222 33244
24
24
222 32 2 334
442
4432
43
4444332 322 3444
32 324
22 34
224
3 34 4
322 332
4
344
24
3222
44 422 3
422 32 2
422 32 332
43
4
224
244 4
242222 34
3 322
4332 32
444443
4432
4432
42 3
422
4432 322222 322 3
43
43 33
44222 22 322
434
244
2 322 324
3224
342 33
4222 122
433
432
43
4 4
34
2 34
32 34
22433332
4
3324
3222
4
2224
22 32 34
24
1
4
12 3322
42
424
24 4
2 324
24
33334
34
34
3332
443
433
44412
43
433
433
44442 34
33
42222
432
44
2 342 44
2
443
42 334 44
324
2 34
2 324
322
2 324
344
324
32 2
42 3
42 3
432
42 3
43
42
42 333
4 422 222
4
322 34
323
4
32 32 32 2 32244
34
334
324 442 3
42
2 32 4
322
432 3332 3333
1
334
3244
344
32 322 22 32 32222224
244
224
2 32 3222
432 3
442
443
43
42
42 33
44322 3
42
4433
4
2224
22
44433244333222
4
2 344
43332
42 32
42 333
443
433
442 32 342 32 34
4
333432 2 322
4 42222
4443
4 332 344
2
44344
33 332 4242 3
432 3 3332
222 342 3
14 42 3
42
42
442
44442 32 3
4
344
244
2
44
22244
32 332422 332 33
434
32 2 344
322
34
4 344
34
3224
3324
22 44222 32 323
4
34
22 34
24
22432
4
32 442
44
234331
22
4444442 3
422322 32
42 22 2
44224
3432 3222 3
42 3 324 4
333 34
2 32 44
34
3444
3344
342
42
44
2 3
443222
44442
4
24 444
34
344
2 3
442 22 2 33332
24
22 344
22 314
34 1
4
233333222 332 3
4
2 344
3242
42 3 3
4432
42 4
3222
14
32 33222 3
4422
433
433
43
433
42
44332
42 33 324
32
43332 32 3322 3
432222 3
4
222 332
4422 32 2
24
34244
244
3243
43
4
3
421
324
2 2 324
324
32224
22244
33442 34
2244
2 24 4
4 312 3
4422 33 32
44
2 33332
4
3342 3
4
32 324444
2222 24
322242 34
2 3322 322 32 322
43
1
34444 41444
2 31332
433
4332 334
3244
222 32 3
44332 332
4322 32
43
443
422 32 33 3
444222 34
2
4
244
32 324
2 334
34
332
4
2 334
2 332 32244424
33322
4
32 3322 21
32224424
2 324
24
32 34
2 3324
222444
2 322 342124
34
24
34
343332 33
4 4
333332 34
2 244
24
344
33334
333 34332
224
2242
4332 33
44433
443432 3
4332 33
43333
42222 2
4
344
2 3444
4 322 32 32224
2 2 32 2 32 34
2 2 334 4
3222 334
4432 33312 3
4433
422 3
4332 332
2 4224
21
24
324
2432 33
44 32 3
444
2 322
42 31 32 3
4433
432 32 322 3
4222
42
44322 3
432
443
4 444342 222
42 33
44
2
4
22 344
2244 3322 3
422 32 33
43
432 3
443
44443
4
32 34
2 344
2 3244
34432
44332 3332
443332
4332 2 332
4422 22 324
244
324
224
33334
3
43
433
43
42222
4332 32 32 342
224
22422
422 32 22
42
4
222 32 33
4
22 33333
442 44
3 3443
4422
4442 32 2
4 433
444
3244
34
332224
322 33332 324
332 2 34
3433
42 3
4
324
3224432222
433333
4
322 34
2444
22
2 34
2 332 3 343
442 12
42 3
43
4222
42
44433
4
22 3
44 4
344
243332222
42 3322
432
4442 32
4
24
32
432
4422
4
342
443
444422
4433
42 333
43
4432 333333
43
44
44222224
2 24
42
422
44
344
2 24444
324
34
14
22
4
34
332 2 32 3
412
24
34
34
24
2 3344
32 12 34
32 244
3344
333444
2 32 342 3324
33222 3344
2 3322 224
2 322224 44
2224
2 33344
2 22 32 3322 3
433
432 333
4
322
4
244
332 33334
3 34
2 344
42222 322 32
432
4432 3
4333
43322 332 33
444224
33342
2 323
42
4
3224
2
43322
44333
4222 33
4
34 4
322
24
34
2
43
4
2 333224
343
41
2 3
443
4322 32 33
434
24
32 34434
244
2 34
24
2444
33222
43332
2 2 224
333
4
3244
2 32 32 33344
2 3442 2
1
43222 3
422 32 3
44
3444
332 344
34
32 2 334444
422
43
432 332
2444224
2 2224 4
42222
42 33
442 332444
24
1 1
1
1
11
11
1
1
1
1
2
1
11
11
114
1
11
1
111
1
1
1
1
4
11
2
1
1
1
1
4
1
2
11
1
11
4
1
1
41
1
1
1
1
1
12
1
11 1
1
1
1
114
11
1
1
1
11
11
1
1 11
1
1
1
21
4
1
11
11
1
42
1
11
11
1
1111
2
1
1
2
41
11
1
11 1
1 1
1
11
1
21
1
4
1
1 4
1
11
1
1
1
321
1
11
1
11
1
1
11
1
11
14
1
2
11
11
1
11
1
111
1
1
2
11
1
3
1
311
1
3
1 1
111
1
11
1 1
1
2
144 3
4
1
11
1
1 1
1
11
1
114
24
11
11
21
1
3
1 1
1
1 1 1
1
11
1
111
1
4
1
4
1
111
31
1
111
22
1
1
1
2
1
1
1
1
11
1
4
11
21
1
11
1
1
1
1 111
1
1
1
22
14
1
1
11
1
1
1
1
2
1
2
1
1
1
1
1
1
1
1
11
1
1 2
1
2
11
1
11
1 1
1
1
1
1
11
1
1
1 4
1
1
3
11
1
1
11
11
2
1
42
1
1
1
111
1 1 1
2
11
11
11
1
41
1
1
12
11
1
1
1
1
1
1
11
1
11
1
1
141 2
1
1
11
2
1 11
1
11 1
1
11
41
1
21
1
1 1
1
1
1
11
1
111
1
11
11
11
11
1
14
1
1
1
2
11 11
1
1
11
14 4
14
1
1
1
12
1
1
1
1
11 1
1
11
2
11
11
1
1
11
1
12
1
2
1
1
11
1
1
1
3
1 1
1
1
34
1
1
11
1
1
1
1
4
1 2
111
1
13
1
1
1
1
112 11
11
1
1
1 1
1
11
4
1
1
1
11 11
41
1
1
14 1
1
1
1 1
14
1 2
1
1
4
1
111
2
1 11
4411
1
1
1
1
1
11 1
2
1
1
24
1
1
1
1
1
21
1
1
21
1
1
1
11
2
1
11
1
4
1
1
1
1
VVV
-20 -10 0 10 20
-5
0
5
5 24
54
5 254
5444
5 25 225
44222 2
42
422
4
5 24
544
24
522
42
4 4222
42
422
42
544255
4
25 225
5 5 2
4
2
4445
5444
54
544
225 25
52555 225
445
45
4555 255 22
544
5
4425
42
4444225 255 2444
25 254
55 24
554
2 24 4
255 225
4
244
54
2555
44 4
55 24
55 25 5
455
25 2254
24
554
544 4
54
5555 24
2 255
4225 25
444442
4525
4425
45 2
455
4425 255555 255 2
42
42 22
44555 55 255
424
544
5 255 254
2554
245 22
4555 455
422
425
42
4 4
24
5 2525 24
5542 2225
4
2254
2555
4
5554
55 25 24
54
5
4
55 2255
45
454
54 4
5 25454
22224
24
24
222544
24
22444
254
24
224
224444
5 24
22
45555
425
44
5 255
445
442
45 224 44
254
5 24
5 254
255
5 254
244
255
25 5
45 2
45 2
425
45 2
42
45
45 222
4 455 255
4
255 24
25 2
4
25 25 25 4 25544
24
224
255
445 2
45
5 25 4
255
425 2225 2222
4
224
2544
244
25 255 55 25 25555554
544
554
5 25 2555
425 2
445
442
42
45
45 22
44255 2
45
4422
4
555
455
44422544222555
4
5 2445
22254
5 254
5 22244
24
2244
5 25 255 25 24
4
222425 5 255
5 45555
4442
5 225 244
5
442
4422 225 4
545 2
425 2 2225
555 245 2
55 45 2
45
45
445
44445 22 2
4
244
544
5
44
55544
25 225455 225 22
424255 2
54 255
24
5 244
24
2554
2224
55 44555 25 25 2
4
24
55 24
54
554
22
4
25 445
44
524 2245
5444444
5 2455255 25
4555 5
44554
2425 2555 2
45 2 255 4
222 24
5 25 44
25
2444
2244
255
45
44
5 2
442555
45445
4
54 444
24
244
5 2
445 55 5 222255
4
55 244
55 244
24 2
4
522222555225 2
4
5 244
2545
45 2 2
4425
55
42555
44
25 22555 2
4 455
422
422
52
422
45
44225
45 22 254
25
42225 25 2255 2
425555 2
4
555 225
4455 25 5
54
24545
544
2542
42
4
2
454
254
5 5 254
254
25554
55544
22445 24
5544
5 54
45 22
5 244
55 22 25
44
5 22225
4
2255 2
4
25 254444
5555 54
255545 24
5 2255 255 25 255
42
4
24444 44444
5 22225
422
4225 224
2544
555 25 2
44225 225
4255 25
42
442
455 25 22 2
445555 24
5
4
544
25 254
5 224
24
225
4
5 224
5 225 25544454
22255
4
25 2255 55
25554454
5 254
54
25 245 225
4555
4445 255 2
45555
24
54
24
242225 22
4 4
222225 24
5 544
54
244
22224
222 24225
554
5542
4225 22
45422
442425 2
4225 22
42222
45555 5
4
244
5 2444
5 255 25 25554
5 5 25 5 25 24
55 225 4
2555 224
4425 22225 2
4422
455 2
4225 225
5 4554
51
54
254
5425 22
44 22 2
444
2 255
45 25 25 2
4422
425 25 255 2
4555
45
44255 2
425
442
4 444245 555
45 22
44
5
4
55 244
5544 2255 2
455 25 22
42
425 2
442
44 442
4
25 24
5 244
5 2544
244
2544
225 222544
2 2254
225 5 2254455 55 25
4544
254
55
42222
42
42422
42
45555
4225 25 25 255
554
55555
455 25 55
45
4
555 25 22
4
55 22222
445 44
2 2442
4455
44 45 25 5
4 422
444
25442
422555
4255 22225 25
4225 5 2
42
422
45 2
4
254
2554425555
422222
4
255 24
5444
55
5 24
5 225 2 242
445 25
45 2
42
4555
45
44422
4
55 2
44 4
244
542225555
55 2255
525
4445 25
4
54
25
425
4455
4
255
442
444555
4422
45 222
42
4425 222222
42
4444255554
5 54 4
54
554
42
4 45 5
445
4
254
24
24
55
4
24
2255 25 2
455
54
24
24
54
5 2244
25 45 24
25 544
2245
222444
5 25 245 2254
22555 2244
52255 55
55 25555
4 44555
4
5 22244
5 25 25 2255 2
422
425 222
4
255
4
55 4
225 22224
2 24
5 244
45555 255 25
425
4425 2
4222
52255 225 22
445555
22242
5 252
45
4
2554
5
42255
44222
2555 22
4
24 4
255
54
24
5
42
4
5 222554
242
45
5 2
442
4255 25 22
424
54
25 24424
544
5 24
54
5444
22555
42222
5 5 554
222
4
2544
5 25 25 22245
5 2445 5
2
42555 2
455 25 2
44
2444
225 244
24
25 5 224 444
455
42
425 225
5 444254
5 5554 44
55554
5 2244
5 225444
54
7 7
2
2
41
77
5
3
7
4
5
3
66
42
114
5
66
4
767
7
5
1
5
4
34
5
6
2
5
3
4
7
5
26
4
76
5
4
7
43
1
2
1
6
7
35
6
33 4
7
5
7
144
55
1
5
4
67
33
5
1 11
7
5
1
52
4
7
31
31
6
45
7
12
62
2
1134
5
6
4
5
43
65
3
77 6
3 4
6
24
7
52
3
4
3
5 4
7
41
4
7
4
255
2
13
7
76
1
7
13
4
76
34
4
5
26
26
5
21
7
767
2
1
5
66
1
2
3
237
7
2
6 6
443
2
11
7 6
1
5
145 2
4
5
66
2
7 6
4
51
4
524
54
66
33
53
1
2
4 1
5
75 2
4
66
4
777
3
4
6
5
1
233
22
7
222
55
4
3
1
5
6
2
3
6
51
2
4
51
25
3
72
1
6
2
5 114
6
1
5
55
34
7
3
67
3
3
5
6
5
6
5
5
7
1
6
3
6
2
2
14
2
5 5
7
2
13
7
67
3 4
2
7
4
6
67
3
6
5 4
6
3
2
77
4
5
14
67
5
7
45
1
6
3
777
3 4 1
5
76
77
32
7
43
4
6
25
11
6
1
7
3
6
3
67
1
43
7
1
255 5
5
1
67
5
3 43
6
56 6
4
76
41
3
57
3
7 2
5
3
3
66
2
314
3
57
66
77
33
6
45
2
3
7
5
77 66
7
4
57
35 4
34
1
6
6
25
1
5
3
5
13 1
7
44
5
32
62
3
6
13
5
75
7
2
6
5
77
5
7
3
2
7 6
2
7
24
6
3
67
2
3
3
6
4
5 2
67 6
3
22
7
3
2
3
525 22
66
7
1
76
4
13
4
6
1
2
66 66
43
5
6
24 1
3
2
3 4
24
5 5
4
6
4
1
666
5
6 67
4444
1
7
6
4
6
13 4
5
3
6
54
1
1
6
4
6
51
5
7
57
6
5
6
43
5
7
47
5
4
3
7
5
1
EI
-20 -10 0 10 20
-5
0
5
4 33
43
4 343
4333
4 34 334
33333 3
33
333
3
4 33
433
33
333
33
3 3333
33
333
33
333344
3
34 331
4 4 3
3
3
3313
4333
43
433
334 34
13444 334
334
34
3444 34 4 33
433
4
3334
33
3333334 344 3333
34 343
44 33
443
3 33 3
344 334
3
333
43
3444
33 3
44 33
44 34 4
344
34 3343
33
341
433 3
434444 33
3 344
3334 34
333333
3134
3333
34 3
334
3334 344444 344 3
33
33 33
33444 34 344
333
433
3 344 343
3443
334 33
3444 344
333
334
33
3 3
33
4 3333 33
44333331
3
3343
3444
3
4343
44 34 33
43
3
3
34 3344
34
343
43 3
4 34313
33333
33
33
333433
33
33333
341
33
333
333333
3 33
33
34444
334
33
4 314
333
333
34 333 33
343
1 33
4 343
341
4 343
333
344
34 3
34 3
34 3
334
34 3
33
33
31 333
3 344 344
3
344 33
34 3
3
34 34 34 3 34433
33
333
344
334 3
34
3 31 3
344
334 3334 3333
3
333
3433
333
34 344 44 33 33444443
433
443
4 34 3441
334 3
334
333
33
34
34 33
33344 3
34
1333
3
444
344
33333333333444
3
4 3334
33313
1 313
4 33333
33
3333
4 34 334 34 33
3
333331 4 344
1 34444
3333
4 334 333
4
33333
33 331 3413 3
334 3 3334
344 334 3
11 34 3
33
34
334
33334 33 3
3
333
433
4
33
44433
31 334344 334 33
333
34 4 333 334
33
4 333
33
3443
3333
41 33444 34 333
3
33
44 33
43
44333
3
34 331
33
333
33344
3333334 3
343344 34
34 41 4
33343
3334 3444 3
34 3 241 3
333 33
4 34 33
34
3333
3333
334
34
33
4 3
333444
31334
3
43 333
33
333
4 3
334 44 4 33333
43
44 333
44 333
33 3
3
433333444334 3
3
4 333
3434
34 3 3
3334
11
33444
33
33 33344 3
3344
333
333
43
333
34
33334
34 33 343
34
33334 34 3334 3
334444 3
3
444 334
3344 34 3
43
33431
433
3433
33
3
3
341
333
4 4 343
343
34343
44433
33333 33
4433
4 33
34 33
4 333
34 33 34
33
4 33334
3
3314 3
3
34 343333
4444 43
344431 33
4 3344 344 33 344
33
1
33333 33333
4 33334
333
3334 333
3433
441 34 3
33334 334
3344 34
33
333
344 34 33 3
334444 33
1
3
433
34 343
4 333
33
334
3
4 333
4 334 34433343
33344
3
34 3344 41
34443333
4 343
43
34 331 334
3444
3334 334 3
34441
33
43
33
333334 33
3 3
333334 33
4 433
43
333
33333
333 33334
443
4433
3334 33
34333
333331 3
3334 33
33333
34441 4
3
333
3 3333
4 344 34 34443
4 3 34 3 34 33
4 4 331 3
3414 333
3333 33334 3
3333
344 3
3333 334
4 3443
43
43
343
4334 33
33 33 3
333
3 344
34 34 34 3
3333
334 34 344 3
3134
34
33344 3
334
333
3 333334 344
34 33
33
4
3
33 333
4433 3344 3
344 34 33
33
334 3
333
33333
3
34 33
4 333
4 3433
33334
33334 3334
333334
3334 3 334
3341 43 343
413
333
44
33333
33
33
333
33
34444
3334 34 34 341
443
44144
344 34 31
34
3
444 34 33
3
44 33333
334 33
3 3313
3334
1334 34 3
3 333
333
3333
33
334443
344 33334 343
331 4 33
3333
34 3
3
313
3443334444
333333
3
344 33
3333
34
4 33
4 331 3 333
334 34
34 3
33
3444
34
33333
3
44 3
33 3
333
433334444
43 3334
134
3314 34
3
43
33
334
3344
3
314
313
333443
3333
34 333
33
3334 333333
33
3333334443
4 333
43
413
33
334 4
334
3
343
33
33
44
3
33
334 4 34 3
313
43
33
33
43
4 3333
34 34 33
34 333
3331
333333
4 34 334 3313
33444 3333
43311 44
41 34414
3 33444
3
4 33333
4 34 34 3344 3
333
334 333
3
344
3
413
334 33333
3 33
4 333
34434 341 34
334
3334 3
3333
43344 334 33
331441
33333
4 343
34
3
3443
4
33344
33333
3444 33
3
31 3
341
43
33
4
13
3
4 333443
333
31
4 3
333
3344 34 33
333
43
34 33333
433
4 33
43
4333
33441
33333
4 4 343
333
3
3433
4 34 34 33331
4 3331 4
3
33444 3
344 34 3
33
3333
334 331
33
31 4 333333
344
33
334 334
4 333343
1 4443 3
34444
34 33
334 334
3334
3
2 2
3
3
22
22
2
1
2
1
3
1
22
12
223
2
22
1
222
2
1
3
2
3
11
1
2
2
1
1
3
2
4
22
3
22
4
2
2
31
2
2
2
2
2
14
2
11 1
2
4
2
213
41
2
2
1
22
11
3
2 22
2
1
2
42
3
2
12
12
2
31
2
22
22
2
2212
3
2
1
4
31
22
1
22 2
1 2
2
33
2
42
1
3
1
1 3
2
32
2
2
3
342
3
21
2
12
2
2
21
3
22
13
3
1
22
32
2
32
2
222
2
2
1
22
3
3
1
311
2
3
2 2
211
2
22
1 2
2
3
234 3
3
1
22
3
2 2
1
12
2
223
43
22
11
11
2
3
1 2
1
2 1 2
1
22
2
222
1
3
2
4
2
311
33
2
222
43
2
1
2
4
2
3
1
2
13
3
3
42
32
1
12
2
2
3
1 221
2
2
2
11
13
2
1
22
1
1
1
2
4
2
1
2
2
2
2
1
2
3
2
21
2
1 4
2
3
21
2
22
1 1
3
2
1
2
21
1
2
1 3
2
1
3
22
2
4
22
22
4
2
31
2
2
1
222
1 1 2
1
22
22
12
2
31
1
2
24
22
2
2
2
1
2
1
22
2
11
1
2
341 3
2
2
21
4
1 11
2
22 2
2
22
32
1
11
1
2 2
1
1
1
22
2
121
1
22
22
22
11
2
31
2
1
2
4
22 22
2
2
12
14 3
13
2
2
2
24
2
1
1
4
21 2
2
31
4
13
23
1
2
21
1
11
2
3
2
1
22
1
2
1
3
2 2
3
2
31
2
1
22
2
1
1
2
3
1 3
22 2
1
33
2
1
2
1
234 23
22
2
2
22
3
21
3
2
2
2
22 22
31
1
2
23 2
1
2
1 1
33
1 4
2
2
3
2
222
4
2 22
3311
2
2
2
1
2
21 1
4
1
2
43
2
2
2
2
2
42
1
2
32
2
1
2
11
1
2
31
1
3
1
2
1
2
VI
-20 -10 0 10 20
-5
0
5
4 63
43
4 643
4333
4 64 664
33666 6
36
366
3
4 63
433
63
466
36
3 3666
36
366
36
333644
3
44 6644 4 6
3
6
3333
4333
43
433
664 643
6444 66433
43
43
444 644 663
334
3364
36
3333664 644 6333
64 643
44 63
443
6 63 3
644 664
3
633
43
6444
33 3
44 63
44 64 4
344 64 664
36
3
443
433 3
43
4444 63
6 644
3664 64
333336
3364
3364
34 6
344
3364 644444 644 6
36
36 66
33444 44 644
363
433
4 644 643
6443
634 66
3444 644
366
364
36
3 3
63
4 6364 63
4436 6664
3
6643
6444
3
4643
44 64 63
43
4
3
44 6644
34
343
43
3
4 64343
66663
63
63
666433
63
66333
643
63
663
663333
4 63
66
34444
364
33
4 634
334
336
34 663 33
643
4 63
4 643
644
4 643
633
643 64 4
34 6
34 6
364
34 6
36
34
34 666
3 344 444
3
644 63
64 6
3
64 64 64 6 64433
63
663
643
334 6
34
4 64 3
644
364 6664 6666
3
663
6433
633
64 644 44 64 64444443
433
443
4 64 6444
364 6
334
336
36
34
34 66
33644 6
34
3366
3
4443
44
33366433666444
3
4 6333
66643
4 643
4 66633
63
6633
4 64 634 64 63
3
666364 4 644
3 34444
3336
3 664 633
4
336
3366 664 3
434 6
364 6 6664
444 63
4 633 34 6
34
34
334
33334 64 6
3
633
433
4
33
44433
64 664344 664 66
363644 6
33 644
63
3 63
3
63
6443
6643
44 33444 64 64 6
3
63
44 63
43
443
66
3
64 33
433
463
66644
3333334 6344644 64
3444 4
33443
6364 6444 6
34 6 643 3
666 63
4 64 33
63
6333
6633
634
34
33
4 6
336444
33334
3
43 333
63
633
4 6
334 44 4 666644
3
44 633
44 663
63 6
3
466666444 664 6
3
4 633
6434
34 6 6
3364
34
36444
33
64 66444 6
3 344
366
366
3 63
663
433
6643
4 66 643
64
36664 64 6644 6
364444 6
3
444 664
3344 64 4
43
63433
433
6436
36
3
6
343
663
4 4 643
643
64443
44433
66334 63
4433
4 43
33 66
4 633
44 66 64
33
4 66664
3
6634 6
3
64 643333
4444 43
644434 63
4 6644 644 64 644
36
3
63333 36333
4 63664
366
3664 663
6433
444 64 6
33664 664
3644 64
36
336
344 64 66 6
333444 63
4
3
433
64 643
4 663
63
664
3
4 663
4 664 64433343
66644
3
64 6644 43
64443343
4 64
343
64 634 664
3444
3334 644 6
34443
63
43
63
636664 66
3 3
666664 63
4 433
43
633
66663
666 63664
443
4436
3664 66
33366
33636
4 63
664 663
66663
4444 43
633
4 6333
3 644 64 64443
4 4 64 4 64 63
44 663 3
6444 663
3364 66664 6
3366
344 6
3664 664
4 3443
43
43
643
4364 66
336
4 6
333
4 644
34 64 64 6
3366
364 64 644 6
3444
34
33644 6
364
336
3 333634 444
34 66
33
4
3
44 633
4433 6644 6
344 64 66
36
364 6
336
33 336
3
64 63
4 633
4 6433
633
6433
664 666433
6 6643
664 4 6643344 44 64
3433
643
443
66663
6
36366
36
34444
3664 64 64 634
443
44344
344 64 44
34
3
444 64 66
3
44 66666
334 33
6 6336
3344
33 34 64 6
3 366
333
64336
366444
3644 66664 64
3664 4 6
36
366
34 6
3
64 36443364444
366666
3
644 63
4333
44
4 63
4 664 6 636
334
643
4 63
63
4443
4333
663
44 6
33 3
633
436664444
34 6644
364
3334 64
3
43
64
364
3344
3
634
336
333344
3366
34 666
36
3364 666666
36
3333444443
4 43 3
43
443
36
3 34 4
333
3
64363
63
44
3
63
6644 64 6
334
43
63
63
43
4 6633
64 34 63
64 433
6633
666333
4 64 634 6643
66444 6633
4 6644 443
4 644443 33
4443
4 66633
4 44 64 6644 6
366
364 666
3
644
3
43 3
664 66663
6 63
4 633
34444 644 64
364
336
4 63
6663
6644 664 66
333443
66636
4 646
34
3
6443
4
36644
33666
3444 66
3
63 3
6444
363
4
36
3
4 666443
636
344 6
336
3644 64 66
363
43
64 63363
433
4 63
43
4333
66444
36664
4 4 443
666
3
6433
4 64 64 66633
4 6334 4
4
36444 6
344 64 6
33
6333
664 633
63
64 4 663 333
344
36
364 664
4 333443
4 4443 33
44443
4 6633
4 664333
43
7 2
6
6
51
77
4
5
7
5
4
5
22
36
113
4
22
5
7227
7
3
4
3
55
4
2
2
4
5
3
7
4
22
3
7 2
3
5
7
34
1
6
1
2
7
44
2
55 5
7
4
7
133
44
1
4
5
27
55
4
1 11
2
7
1
46
3
7
51
51
2
34
7
16
26
2
1155
4
2
5
4
35
24
5
77 2
5 5
2
36
7
46
5
3
5
3 3
7
3 15
7
3
644
6
15
7
72
1
7
15
3
72
53
3
4
22
62
4
33
7
722
6
3
4
22
3
6
5
647
7
6
2 2
555
6
15
7 2
5
4
133 6
3
4
22
6
7 2
5
41
5
463
43
22
45
44
1
6
5 5
4
77 6
5
22
5
277
5
3
2
3
1
645
66
2
666
44
5
4
1
4
2
6
5
2
43
6
3
41
47
5
72
1
2
6
3 113
2
1
7
44
53
7
5
27
4
5
7
2
4
2
4
4
7
1
2
5
2
6
4
15
4
7 4
7
6
15
7
22
5 5
4
2
5
2
67
4
2
4 3
2
5
6
27
5
4
15
22
4
7
34
1
2
5
777
55 1
4
72
77
56
7
35
5
2
64
11
2
1
7
5
2
5
27
1
55
7
5
634 4
4
1
27
4
5 55
2
42 2
5
72
31
5
47
5
7 6
4
5
4
22
6
415
5
77
22
77
54
2
33
4
5
7
4
77 22
7
5
47
53 3
53
1
6
2
64
1
7
5
4
15 1
7
35
4
53
24
5
2
15
4
74
7
4
2
4
77
4
7
5
6
2 2
3
7
63
2
5
27
6
5
4
2
3
7 6
27 2
5
66
7
5
4
5
464 66
22
7
1
72
6
15
3
2
5
4
22 22
35
7
2
63 1
5
4
5 5
33
7 4
5
2
3
1
222
4
2 27
3353 1
2
2
5
2
15 5
4
5
2
43
1
1
2
5
2
41
4
7
47
2
4
2
35
4
7
67
7
3
5
7
4
1
EVE
-20 -10 0 10 20
-5
0
5
3 54
34
3 534
3444
3 53 553
44555 5
45
455
4
3 54
344
54
355
45
4 4555
45
455
45
444533
4
33 5513 3 5
4
5
4444
3444
34
344
553 534
5333 55344
34
34
333 53 3 554
443
4453
45
4444553 533 5444
53 534
33 54
334
5 54 4
533 553
4
544
34
5333
44 4
33 54
33 53 3
433 53 553
45
4
334
344 4
343333 54
5 533
4553 53
444445
4453
4453
43 5
433
4453 533333 533 5
45
45 55
44333 33 533
454
344
3 533 534
5334
543 55
4333 333
455
453
45
4 6
54
3 5453 54
33455551
4
5534
5333
4
3334
33 53 54
34
3
4
33 5533
43
434
34
4
3 53414
55554
54
54
555344
54
55444
534
54
554
554444
3 54
55
43333
453
44
3 543
443
445
43 554 44
534
3 54
3 534
533
3 534
544
534 53 3
43 5
43 5
453
43 5
45
43
43 555
4 433 333
4
533 54
53 5
4
53 53 53 3 53344
54
554
534
443 5
43
3 51 4
533
453 5553 5555
6
554
5344
544
53 533 33 53 53333334
344
334
3 53 5331
453 5
443
445
45
43
43 55
44533 5
43
4455
4
3334
33
44455344555333
4
3 5444
55534
3 534
3 55544
54
5544
3 53 543 53 54
4
555453 3 533
4 43333
4445
4 553 544
3
44344
55 553 4343 5
453 5 5553
333 54
3 544 43 5
43
43
443
44443 53 5
4
544
344
3
44
33344
53 553433 553 55
454
53 3 544 533
54
4 544
54
5334
5534
31 44333 53 535
4
54
33 54
34
33453
4
53 44
144
354
55633
4444443 5
433533 53
43 33 3
44334
5453 5333 5
43 5 534 4
555 54
3 53 44
54
5444
5544
543
43
44
3 5
445333
44443
4
34 444
54
544
3 5
443 33 3 55553
34
33 544
33 544
54 5
4
355555333 553 5
4
3 544
5343
43 5 5
4453
43
45333
64
53 55333 5
4433
455
455
4 54
554
344
5534
3 55 534
51
45553 53 5533 5
453333 5
4
333 553
4433 53 3
34
54344
344
5345
45
4
5
434
534
3 3 534
534
53334
33344
55443 54
3344
3 34
64 55
3 544
33 55 53
44
3 55553
4
5543 5
4
53 534444
3333 34
533343 54
3 5533 533 53 533
45
6
54444 44444
3 56553
455
4553 554
5344
333 53 5
44553 553
4533 53
45
445
433 53 55 5
444333 54
1
4
344
53 534
3 554
54
553
4
3 554
3 553 53344434
55533
4
53 5533 34
53334434
3 53
434
53 543 553
4333
4443 533 5
43334
54
34
54
545553 55
4 4
555553 54
3 344
34
544
55554
555 54553
334
3343
4553 55
44455
44545
3 54
553 554
55554
3333 34
544
3 5444
4 533 53 53334
3 3 53 3 53 54
3 3 554 4
5333 554
4453 55553 5
4455
433 5
4553 553
3 4334
36
34
534
3453 55
445
3 5
444
3 533
43 51 33 5
4455
453 53 533 5
4333
43
44533 5
453
445
4 444543 333
43 55
44
3
4
33 544
3344 5533 5
433 53 55
45
453 5
445
44445
4
53 54
3 544
3 5344
54453
44553 5553
445553
4553 3 553
4433 33 534
344
534
334
53554
5
45
455
45
43333
4553 53 53 543
334
33433
433 53 33
43
4
333 53 53
4
33 55555
443 44
5 5445
4433
4443 53 3
4 455
444
5344
54
553334
533 55553 534
553 3 54
5455
43 5
4
53 45334453333
455555
4
533 54
3444
33
3 54
3 551 5 545
443
334
3 54
54
3334
3444
554
33 5
44 4
544
345553333
43 5533
453
4443 53
4
34
53
453
4433
4
543
445
444433
4455
43 555
45
4453 555555
45
4444333334
3 344
34
334
45
443 3
444
4
53454
54
33
4
54
553 3 53 5
413
34
54
54
34
3 5544
53 43 54
53 344
5544
555444
3 53 543 5534
55333 5344
3 5533 334
3 533334 44
3334
3 55544
3 33 53 5533 5
455
453 555
4
533
4
344
553 55554
5 54
3 544
43333 533 53
453
445
3 54
5554
5533 553 55
444334
55543
3 535
43
4
5334
3
45533
44555
4333 55
4
54 4
533 3
454
3
45
4
3 555334
545
413 5
445
4533 53 55
454
34
53 54454
344
3 54
34
3444
55333
45553
3 3 334
555
4
5344
3 53 53 55544
3 5443 3
3
45333 5
433 53 5
44
5444
553 544
54
53 3 554444
433
45
453 553
3 444334
3 3334 44
33334
3 5544
3 553444
34
1 2
2
6
66
12
2
1
1
6
3
6
22
66
664
1
22
6
2221
1
6
2
4
16
3
2
2
1
6
4
2
3
22
6
1 2
4
6
1
41
6
2
6
2
2
13
2
61 6
1
1
1
664
11
6
1
6
21
11
3
6 66
2
1
6
32
4
1
16
16
2
41
2
66
22
2
6616
3
2
6
3
46
22
1
11 2
6 6
2
63
1
32
6
4
1
1 4
1
6 66
1
6
531
2
61
2
12
6
1
66
6
12
14
6
3
22
22
1
66
1
222
2
6
3
22
6
5
6
511
1
5
2 2
661
2
66
1 2
6
3
643 5
4
1
22
5
2 2
6
16
6
364
34
22
16
11
6
5
6 6
1
1 1 2
6
22
6
211
1
4
2
4
6
511
52
2
626
33
6
1
6
3
2
2
1
2
16
5
4
36
31
6
12
6
2
2
1 666
2
6
1
11
64
1
6
21
1
6
1
2
3
2
1
1
2
6
2
6
2
6
2
66
2
1 3
1
3
66
1
22
6 6
2
2
6
2
21
1
2
1 4
2
6
5
22
6
1
66
22
3
1
43
6
2
1
111
16 6
1
12
11
16
1
41
6
2
23
66
2
6
1
1
2
1
21
6
66
1
6
641 3
2
6
21
3
1 61
2
22 2
6
22
46
1
11
6
1 2
1
6
1
22
2
166
1
11
22
11
11
2
64
2
6
2
3
11 22
2
6
11
14 4
64
6
2
2
23
6
1
1
1
61 6
2
66
3
66
22
6
2
66
1
11
1
3
2
1
11
1
1
6
5
2 2
6
2
54
2
1
22
2
6
1
2
4
1 3
22 2
1
65
1
1
2
6
233 63
22
1
6
22
3
66
4
2
6
2
22 22
41
1
2
64 6
6
2
1 6
64
1 3
6
2
4
6
222
3
2 21
4466 6
2
2
6
2
61 6
3
6
2
34
6
6
2
6
2
36
1
1
31
2
3
2
46
1
1
31
1
6
6
2
1
6
VVE
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7-3.4
-3.3
-3.2
-3.1
-3
-2.9
-2.8x 10
4
cluster numbers
BIC
va
lue
s
VVV
EI
VI
EVE
VVE
Fig. 21. Data set with 600 noisy points and different clustering results and BIC values for Example 7.
5. Conclusions
Using the fuzzy classification maximum likelihood (FCML) with five eigenvalue
decomposition models and BIC model selection, we propose a robust FCML clustering
framework. The proposed clustering method achieves robust clustering results, with
robustness to different cluster volumes and also cluster number and with no parameter
initialization problem. Several simulated data and real data sets are used to demonstrate
the effectiveness and usefulness of the proposed model. A comparison of the proposed
method with the FCM, PFCM, and GK algorithms shows that the proposed method
produces better results for the simulated and real data sets. In fact, the FCM algorithm
can be derived through the FCML, which means that the proposed model can be seen as a
probability model for the FCM. Further study of other relationships and properties, such
as stochastic convergence and limiting theorems for the FCM algorithm, are worthwhile,
using the proposed model. New fuzzy clustering algorithms may also be developed using
the proposed robust FCML clustering framework.
A Robust Fuzzy Classification Maximum Likelihood Clustering Framework 775
Acknowledgements
The authors are grateful to the anonymous referees for their constructive comments and
suggestions to improve the presentation of the paper. This work was supported in part by
the National Science Council of Taiwan, under Grant NSC101-2118-M-033-002-MY2.
References
1. J. A. Hartigan, Clustering Algorithms (Wiley, New York, 1975).
2. L. Kaufman and P. J. Rousseeuw, Finding Groups in Data: An Introduction to Cluster
Analysis (Wiley, New York, 1990).
3. R. Duda and P. Hart, Pattern Classification and Scene Analysis (Wiley, New York, 1973).
4. A. K. Jain and R. C. Dubes, Algorithms for Clustering Data (Prentice-Hall, New Jersey,
1988).
5. G. J. McLachlan and K. E. Basford, Mixture Models: Inference and Applications to
Clustering (Marcel Dekker, New York, 1988).
6. S. Zhong and J. Ghosh, A unified framework for model-based clustering, Journal of Machine
Learning Research 4 (2004) 1001−1037.
7. J. MacQueen, Some methods for classification and analysis of multivariate observations,
Proc. 5th Berkeley Symposium on Mathematical Statistics and Probability, Vol. 1, University
of California Press, Berkley, 1967, pp. 281−297.
8. C. Fraley and A.E. Raftery, How many clusters? Which clustering method? Answers via
model-based cluster analysis, Computer Journal 41 (1998) 586−588.
9. L. A. Garcia-Escudero and A. Gordaliza, Robustness properties of k means and trimmed k
means, Journal of the American Statistical Association 94 (1999) 956−969.
10. J. A. Guesta-Albertos, A. Gordaliza and C. Matran, Trimmed k-means: an attempt to
robustify quantizers, The Annals of Statistics 25 (1997) 553−576.
11. J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms (Plenum Press,
New York, 1981.
12. F. Hoppner, F. Klawonn, R. Kruse and T. Runkler, Fuzzy Cluster Analysis: Methods for
Classification Data Analysis and Image Recognition (Wiley, New York, 1999).
13. M. S. Yang, A survey of fuzzy clustering, Mathematical and Computer Modeling 18 (1993)
1−16.
14. Y. Cheng, Mean shift, mode seeking, and clustering, IEEE Trans. Pattern Analysis and
Machine Intelligence 17 (1995) 790−799.
15. K. Fukunaga and L.D. Hostetler, The estimation of the gradient of a density function with
applications in pattern recognition, IEEE Trans. Information Theory 21 (1975) 32−40.
16. K. L. Wu and M. S. Yang, Mean shift-based clustering. Pattern Recognition 40 (2007)
3035−3052.
17. R. Krishnapuram and J. M. Keller, A possibilistic approach to clustering, IEEE Trans. Fuzzy
Systems 1 (1993) 98−110.
18. N. R. Pal, K. Pal, J. M. Keller and J. C. Bezdek, A possibilistic fuzzy c-means clustering
algorithm, IEEE Trans. Fuzzy Systems 13 (2005) 517−530.
19. M. S. Yang and K. L. Wu, Unsupervised possibilistic clustering, Pattern Recognition 39
(2006) 5−21.
20. A. P. Dempster, N. M. Laird and D. B. Rubin, Maximum likelihood from incomplete data via
the EM algorithm (with discussion), J. Roy. Stat. Soc., Ser. B 39 (1977) 1−38.
21. G. J. McLachlan and T. Krishnan, The EM Algorithm and Extensions (Wiley, New York
1997).
22. G. Bryant and J. A. Williamson, Asymptotic behaviour of classification maximum likelihood
estimates, Biometrika 65 (1978) 273−281.
M.-S. Yang, C.-Y. Lin & Y.-C. Tian 776
23. A. J. Scott and M. J. Symons, Clustering methods based on likelihood ratio criteria,
Biometrics 27 (1971) 387–397.
24. M. J. Symons, Clustering criteria and multivariate normal mixtures, Biometrics 37 (1981)
35–43.
25. M. S. Yang, On a class of fuzzy classification maximum likelihood procedures, Fuzzy Sets
and Systems 57 (1993) 365-375.
26. J. D. Banfield and A. E. Raftery, Model-based Gaussian and non-Gaussian clustering,
Biometrics 49 (1993) 803−821.
27. C. Fraley and A. E. Raftery, How many clusters? Which clustering method? Answers via
model-based cluster analysis, Computer Journal 41 (1998) 586−588.
28. C. Fraley and A.E. Raftery, Model-based clustering, discriminant analysis, and density
estimation, Journal of the American Statistical Association 97 (2002) 611−631.
29. W. L. Hung, Y. C. Chang and S. C. Chuang, Fuzzy classification maximum likelihood
algorithms for mixed-Weibull distributions, Soft Computing 12 (2008) 1013−1018.
30. C. Y. Lin and C. B. Chen, An invisible hybrid color image system using spread vector
quantization neural networks with penalized FCM, Pattern Recognition 40 (2007)
1685−1694.
31. J. S. Lin, K. S. Cheng and C. W. Mao, Segmentation of multispectral magnetic resonance
image using penalized fuzzy competitive learning network, Computers and Biomedical
Research 29 (1996) 314−326.
32. S. H. Liu and J. S. Lin, Vector quantization in DCT domain using fuzzy possibilistic c-means
based on penalized and compensated constraints, Pattern Recognition 35 (2002) 2201−2211.
33. M. S. Yang and N. Y. Yu, Estimation of parameters in latent class models using fuzzy
clustering algorithms, European Journal of Operational Research 160 (2005) 515−531.
34. L. A. Zadeh, Fuzzy sets, Information and Control 8 (1965) 338−356.
35. E. Ruspini, A new approach to clustering, Information and Control 15 (1969) 22−32.
36. J. H. Ward, Hierarchical groupings to optimize an objective function, Journal of the
American Statistical Association 58 (1963) 236−244.
37. H. P. Friedman and J. Rubin, On some invariant criteria for grouping data, Journal of the
American Statistical Association 62 (1967) 1159−1178.
38. C. L. Blake and C. J. Merz, UCI repository of machine learning databases, a huge collection
of artificial and real-world data sets, 1998. Website http://archive.ics.uci.edu/ml/datasets.html.
39. G. M. Reaven and R. G. Miller, An attempt to define the nature of chemical diabetes using a
multidimensional analysis, Diabetologia 16 (1979) 17−24.
40. W. H. Wolberg and O. L. Mangasarian, Multisurface method of pattern separation for
medical diagnosis applied to breast cytology, Proc. National Academy of Sciences 87 (1990)
9193−9196.
41. S. Chatzis and T. Varvarigou, Robust fuzzy clustering using mixtures of Student’s-t
distributions, Pattern Recognition Letters 29 (2008) 1901−1905.
42. S. Chatzis and T. Varvarigou, Factor analysis latent subspace modeling and robust fuzzy
clustering using t-distributions, IEEE Transactions on Fuzzy Systems 17 (2009) 505−517.
43. D. Peel and G. J. McLachlan, Robust mixture modelling using the t distribution, Statistics
and Computing 10 (2000) 339–348.
44. F. Greselin and S. Ingrassia, Constrained monotone EM algorithms for mixtures of
multivariate t distributions, Statistics and Computing 20 (2010) 9−22.