Data Mining Techniques for Segmentation - Data Mining Techniques in CRM: Inside Customer Segmentation - page 121

High roaming usage leads to the rightmost node 6 which predominantly contains
members of cluster 4. On the other side of this branch, high voice and relatively
lower roaming usage results in node 5 which is dominated by cluster 2. Customers
with relatively lower voice usage are grouped further according to their SMS
usage. Customers with high SMS usage land on node 4, in which cluster 3 is the
modal category. On the left side of the branch, customers with lower SMS (and
voice) usage end up in node 3 which mainly consists of members of cluster 6.
Although additional partitions not shown here further refine the separation and
outline the differentiating characteristics of all clusters, the fist two splits of the
model have already started to portray the profile of specific clusters (clusters 4,
2, 3, and 6).
As mentioned above, behind each terminal node there is a relevant classifi-
cation rule. In Figure 3.28 we present some of these rules which, as explained
earlier, can score and assign new records to the revealed clusters.
Each rule is presented with the relative:
• support - the number of customers at the respective terminal node; and
• confidence - the proportion of the modal category, in this example cluster, in
the respective terminal node.
By examining the rules subset presented in the screenshot of Figure 3.28, we
can see that cluster 3 is mainly associated with increased SMS usage, cluster 4 with
high roaming usage, and cluster 6 with relatively low usage of all services.
THE ADVANTAGES OF USING DECISION TREES FOR CLASSIFICATION
MODELING
Decision trees are among the most commonly used classification techniques since
they present significant advantages:
• They generate simple, straightforward, and understandable rules which provide
insight into the way that predictors are associated with the output.
• They are fast and scalable and they can efficiently handle a large number of
records and predictors.
• They integrate a field screening process which permits them to narrow the set of
predictors by filtering out the irrelevant ones. The best predictors are selected
for each partition. Predictors with marginal relevance to the output are filtered
out from the model training procedure.
• They are not affected by the possible correlation between predictors and the
presence of multicollinearity, a situation that can seriously affect the stability of
other classification models.