rmany
n
lection.
stem to
Modern
provide
nethods
-means
nfusion
e using
ries or
of the
in the
Istering
Izzy C-
(plored
in the
which
ership
vector
ick up
cation.
ype of
nizing
ing, is
Is. The
rough
et of n
centre
ilarity,
jective
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
function that should be minimized, when the Euclidean distance
is selected as dissimilarity measures can be described as:
7-5 X
i=l \ k.x, EG;
ju —ef
NK i
er | is the objective function within group i,
EK i
where €
k.x; €0;
2, :
-c| is a chosen distance measure between a data
point x, and the cluster centre C;.
The partitioned groups are typically defined by a C X71 binary
membership matrix U, where the element 4; is 1 if the jth
data point x; belongs to group i, and 0 otherwise. That means:
if EN
ij | | 0 otherwise (2)
<|x, Ze for each kzi,
u
The algorithm of K-means clustering is consists of the
following steps:
Step 1: Initialize the cluster centres C Em, c This is
typically achieved by randomly selecting c points from among
all of the data.
Step 2: Determine the membership matrix U according to
Equation (2).
Step 3: Compute the objective function according to Equation
(1). Stop if either it is below a certain tolerance value or its
improvement over previous iteration is below a certain
threshold.
Step 4: Update the cluster centres. Go to step 2.
The algorithm is significantly sensitive to the initial randomly
selected cluster centres. The K-means algorithm can be
employed multiple times to reduce this effect. More details can
be found in (Jang 1997).
2.20 Fuzzy C-means clustering algorithm:
Fuzzy C-means clustering (FCM), also known as fuzzy
ISODATA, is a data clustering algorithm in which each data
point belongs to a cluster to a degree specified by its
membership grade. Bezdek (1981, 1987) proposed this
algorithm as an alternative to earlier (hard) K-means clustering.
FCM partitions a collection of n vector x, /—1,...,n into c
fuzzy groups, and finds a cluster centre in each group such that
an objective function of a dissimilarity measure is minimized.
The major difference between FCM and K-means is that FCM
employs fuzzy partitioning such that a given data point can
belong to several groups with the degree of belongingness
specified by membership grades between 0 and 1. In FCM the
membership matrix U is allowed to have not only 0 and 1 but
also the elements with any values between 0 and 1.
$ u; = Ll, Vj zl,
i=l
(3)
The objective function for FCM is then a generalization of
Equation (1) as following:
39
T
(4)
J(U, 6 e.
=) Yu ds, - C
ib j=l
,C.)= S
izl
Where uj; is between 0 and 1; c; is the cluster centre of fuzzy
group i. Fuzzy partitioning is carried out through an iterative
optimization of the objective function shown above, with the
update of membership u;; and the cluster centres c; by:
3S Vs (5)
e=—,
i n m
Y tg
| (6)
Us = 27
3 /(m- n
(=e)
Ze J, el
The FCM clustering algorithm is composed of the following
steps:
Step I: Initialized the membership matrix U with random values
between 0 and 1 such that te constraints in Equation (3) are
satisfied.
ii.
Step 2: Calculate c fuzzy cluster centres C; , using
Equation (5).
Step 3: Compute the cost function (objective function)
according to Equation (4). Stop if either it is below a certain
tolerance value or its improvement over previous iteration is
below a certain threshold.
Step 4: Compute a new U using Equation (6). Go to step 2
The cluster centres also can be initialized firstly and then
iterative process carried out. (Jang 1997) provides a detailed
behaviour of fuzzy c-means clustering including its variants and
convergence properties.
2.3 Competitive Learning Networks
Competitive leaning networks are an unsupervised learning
method which is based on the neural network concept. But
unlike other neural learning algorithms in unsupervised
classification no external teacher is available, thus only input
vectors can be used for learning. In the following a competitive
learning algorithm is described. Competitive learning is a self-
organizing property of the neural network.
The method updates weights only on the basis of the input
patterns. Figure (1) presents an example. All input units / re
connected to all output units j with weight w;. The number of
inputs is the input dimension, while the number of outputs is
equal to the number of clusters. A cluster's centre position is
specified by the weight vector connected to the corresponding
output unit.
W411 1] >
X1
20>
X2 :
3i >
X3
Was 41>
Figure 1: Competitive learning network