Data Mining Techniques for Segmentation - Data Mining Techniques in CRM: Inside Customer Segmentation - page 131

is characterized by its centroid or cluster center. This is a virtual ''prototype
record,'' the vector of the means of all clustering inputs. It is the central point of
the cluster. Cluster centers can be considered as themost typical member cases of
each cluster and therefore are usually examined and explored for cluster profiling.
• Apart from simple tables and charts of descriptive statistics, revealed clusters
can also be profiled through the use of classification modeling techniques which
examine the association of the clusters with the fields of interest. Decision trees,
due to their transparency and the intuitive form of their results, are commonly
applied to explore the structure of the clusters.

Table 3.15 summarizes and compares the characteristics of the clustering
techniques presented in the previous sections, namely the K-means, TwoStep, and
Kohonen network techniques.

Table 3.15 Comparative table of clustering techniques.

K-means TwoStep Kohonen

network/SOM

Methodology
description

Iterative proce-
dure based
on a selected
(typically
Euclidean)
distance
measure

In the first phase records,
through a single data
pass, are grouped into
pre-clusters. In the sec-
ond phase pre-clusters
are further grouped
into the final clusters
through hierarchical
clustering

Based on neural
networks. Clus-
ters are spatially
arranged in a grid
map with dis-
tances indicating
their similarities

Handling of
categorical
clustering
fields

Yes, through
a recoding
into indicator
fields

Yes Yes, through a

recoding into
indicator fields

Number of
clusters

Analysts specify
in advance
the number
of clusters to
fit; therefore
it requires
multiple runs
and tests

The number of clusters
is automatically deter-
mined according to
specific criteria

Analysts specify the
(maximum) num-
ber of output
neurons. These
neurons indi-
cate the probable
clusters

( continued overleaf )