Soft Learning Vector Quantization and Clustering Algorithms Based on Ordered Weighted Aggre
Soft Learning Vector Quantization and Clustering Algorithms Based on Ordered Weighted
Aggregation Operators
Nicolaos B.Karayiannis,Member,IEEE
Abstract—This paper presents the development and investigates
the properties of ordered weighted learning vector quantization
(LVQ)and clustering algorithms.These algorithms are developed
by using gradient descent to minimize reformulation functions
based on aggregation operators.An axiomatic approach pro-
vides conditions for selecting aggregation operators that lead to
admissible reformulation functions.Minimization of admissible
reformulation functions based on ordered weighted aggregation
operators produces a family of soft LVQ and clustering algo-
rithms,which includes fuzzy LVQ and clustering algorithms as
special cases.The proposed LVQ and clustering algorithms are
used to perform segmentation of magnetic resonance(MR)images
of the brain.The diagnostic value of the segmented MR images
provides the basis for evaluating a variety of ordered weighted
LVQ and clustering algorithms.
Index Terms—Aggregation operator,clustering,image seg-
mentation,learning vector quantization,magnetic resonance
(MR)imaging,ordered weighted aggregation,reformulation,
reformulation function.
I.I NTRODUCTION
C ONSIDER the set fea-
ture vectors from an
feature vectors to
-dimensional Euclidean space
also referred to
as the codebook.Codebook design can be performed by clus-
tering algorithms,which are typically developed to solve a con-
strained minimization problem involving two sets of unknowns,
namely the membership functions that assign feature vectors to
clusters and the prototypes.The solution of this problem is often
determined using alternating optimization[1].These clustering
techniques include the crisp-means(FCM)
[1],generalized fuzzy
Axiom A3:The
function
and
of
represent the membership
functions of the aggregated fuzzy sets and,as such,their values
are typically restricted to the interval[0,1].This is reflected by
the additional axiomatic requirement often imposed on aggre-
gation operators for fuzzy sets,according to which the
function
and
are
not necessarily upper-bounded by one.Moreover,the boundary
conditions mentioned above are satisfied by any
function
must be symmetric under any permutation of its arguments,
that
is,
in
are considered to be equally important
to the aggregation operator,a property that is desirable but
not necessary for the development of soft LVQ and clustering
algorithms.
B.Update Equations
Suppose the development of LVQ algorithms is attempted by
using gradient descent to minimize the function(1).The gra-
dient
of
can be determined
as
(3)
where are the competition functions,defined
as
are determined in terms
of
(6)
The LVQ algorithms described by the update equation(6)re-
duce to iterative clustering algorithms if
[26]
(7)
The closed-form formula that can be used to compute the pro-
totypes at each iteration is obtained by substituting the learning
rates at(7)in the update equation(6)
as
(8)
C.Admissible Reformulation Functions Based on Aggregation
Operators
The search for admissible reformulation functions is based on
the properties of the competition
functions
defined in terms of an aggre-
gation
operator
prototypes Consider the func-
tion in(1).
Then,
is a continuous and differentiable everywhere func-
tion;
2)
KARAYIANNIS:SOFT LEARNING VECTOR QUANTIZATION AND CLUSTERING ALGORITHMS1095
3)
4)
for
used to construct reformulation func-
tions of the form(1)must be an aggregation operator in accor-
dance with the axiomatic requirements A1–A3.Nevertheless,
Theorem1indicates that not all aggregation operators lead to
admissible reformulation functions.The subset of all aggrega-
tion operators that can be used to construct reformulation func-
tions of the form(1)are those satisfying the fourth condition of
Theorem1.
III.R EFORMULATION F UNCTIONS B ASED ON O RDERED
W EIGHTED A GGREGATION O PERATORS
A broad family of aggregation operators is composed of or-
dered weighted operators[21],[27]–[29],[31].Consider the
function
and the
weights
Any function of the form(9)is an admissible refor-
mulation function if the
functions and the
weights
prototypes Consider the
function
(10)
where ordered in ascending
order,that
is,and the
weights
is an admissible reformulation function of the first
(second)kind in accordance with the axiomatic requirements
R1–R3
if are continuous and differentiable every-
where functions
satisfying
and are both
monotonically decreasing(increasing)functions
of
is a monotonically increasing(decreasing)function
of
and
Proof:Proof of this theorem is shown in Appendix B.
It can easily be verified that the
function
is not affected by their permu-
tation.Thus,the function(10)satisfies the auxiliary axiomatic
requirement for aggregation operators in addition to the basic
Axioms A1–A3.
A.The Ordered Weighted Generalized Mean
The development of soft LVQ and clustering algorithms can
be accomplished by considering ordered weighted aggregation
operators of the form(10)corresponding
to
and
Theorem2requires
that be a monotonically
increasing function
of which is true
if
this last inequality is valid
if
corresponding
to
corresponding
to
(11)
with
denotes the ordered
weighted generalized mean
of
(13)
If the ordered weighted generalized mean(12)differs
from the ordered weighted mean(13)unless the weight
vector
then coincides with the general-
ized mean or
unweighted
associated with a weight
vector
1096IEEE TRANSACTIONS ON NEURAL NETWORKS,VOL.11,NO.5,SEPTEMBER2000
that
(14)
By
definition,The values
of
assuming descending ordering
of the
arguments
measures the degree of“orness”of the corresponding ordered
weighted aggregation operator.In the formulation considered
in this paper,it is assumed that the
arguments
are ordered in ascending order.In such a
case,
if
Since
and
decreases from one and approaches zero as the aggregation
operator
to one
can be used to measure the emphasis placed by the
weights
which corresponds
to
defined in(14)takes values in the interval[1/2,1].This can
be proven using the results of the following proposition.
Proposition1:Consider the weight
vector
then
in(14)
that obtained
for
can be obtained by writing
(14)
as
then
According to Proposition
1
(17)
with the equality holding
for
is the percentage of
weights marked for elimination.Since the
weights
denotes the integer part
of
for
,the weights defined in(19)form the
weight
vector
In this case,the ordered weighted generalized mean(12)
coincides with the generalized mean.
For,the weights
defined in(19)form the weight
vector
defined in(14)can be computed
as
,which corresponds
to
while the upper
bound,which corresponds
to
is the percentage of weights marked
for elimination.Since the
weights
For
,then the weights
KARAYIANNIS:SOFT LEARNING VECTOR QUANTIZATION AND CLUSTERING ALGORITHMS1097 defined in(21)form the weight vector
defined in(14)can be computed
as
is not a function of
increases above zero,that takes values
above2/3.The upper bound
,which reduces the ordered weighted generalized mean
(12)to the minimum of
1098IEEE TRANSACTIONS ON NEURAL NETWORKS,VOL.11,NO.5,SEPTEMBER2000 This is the form of the membership functions of the minimum
FCM algorithm[15].
If
then(32)
becomes
then
-partitions[1].In fact,this constraint was
used in the development of the FCM and entropy-constrained
fuzzy clustering algorithms[9],[16].
The
do not satisfy the
constraint
-partitions.Such weight vec-
tors lead to soft LVQ and clustering algorithms.For any set of
weights,the values
of
or
if
Since
and
According to Proposition
2
The constraint(35)satisfied by the membership functions de-
fined in(32)indicates that ordered weighted clustering algo-
rithms can also be derived by using alternating optimization to
solve the constrained minimization
problem
and
the is
a
the definition
of by(41)indicates
that It can also be verified that
the ordered membership functions defined in(41)
satisfy
then(42)
gives
-par-
titions that provided the basis for the development of the FCM
algorithm[1].According to(42),the proposed formulation pro-
duces a variety of algorithms whose behavior and performance
can be tailored to the application at hand by selecting sets of
weights
KARAYIANNIS:SOFT LEARNING VECTOR QUANTIZATION AND CLUSTERING ALGORITHMS1099 which allows them to search for near-optimum partitions in a
subspace of all admissible partitions that is not necessarily re-
stricted to a small neighborhood centered at the initial parti-
tion.Although the initialization of soft LVQ and clustering algo-
rithms has little or no effect on their performance on simple data
sets[13],[14],the application of such algorithms in more chal-
lenging problems indicated that their initialization has a rather
significant effect on their performance.
A.Initialization of Soft LVQ and Clustering Algorithms
Ordered weighted LVQ and clustering algorithms were ini-
tialized in the experiments presented in this paper by a pro-
totype splitting procedure that begins with a single prototype
and designs a codebook containing
is split to
create a codebook of size2containing a new prototype and
an updated version of the original prototype
iterations the
codebook contains the prototypes
formed by the indexes of the
feature vectors represented by the prototype
the prototype is split
if
The new prototype is obtained by moving the orig-
inal prototype toward
the new
prototype
with
is fixed during the
learning process and the learning rates are computed at each
iteration as
and the value of
1100IEEE TRANSACTIONS ON NEURAL NETWORKS,VOL.11,NO.5,SEPTEMBER2000 Segmentation of MR images is formulated to exploit the
differences among local values of the T1,T2,and Flair relax-
ation parameters.The values of these parameters represent the
intensity levels(pixels)of a set of three images,namely the
T1-weighted,T2-weighted,and Flair-weighted images.Let
and be the pixel values of the T1-weighted,
T2-weighted,and Flair-weighted images,respectively,at a
certain location.The relaxation parameter values
and can be combined to form the vector vector
can be represented in the
segmentation process by
and
with
KARAYIANNIS:SOFT LEARNING VECTOR QUANTIZATION AND CLUSTERING ALGORITHMS1101
(a)(b)
(c)(d)
Fig.2.Segmented MR images produced by ordered weighted clustering algorithms corresponding to the weight sets(a)S
1102IEEE TRANSACTIONS ON NEURAL NETWORKS,VOL.11,NO.5,SEPTEMBER
2000
(a)
(b)
(c)
(d)
Fig.3.Segmented MR images produced by ordered weighted LVQ algorithms corresponding to the weight sets (a)S
=5to m
m LV m me m m m LV m m m m
m m m m LV m
m m
m
m N N
LV m mi m m m