Mining Your Own Business in Telecoms Using DB2 Intelligent Miner for Data

58 Mining Your Own Business in Telecoms Using DB2 Intelligent Miner for Data

These two techniques solve the same problems from different views, both can be

used to complement each other and to gain confidence that the segmentation

produced has a valid and optimal result.

4.5.2 Applying the mining techniques

Behavior data is used as active variables to construct the segmentation model

and demographic and additional data are used as only supplementary variables

to describe the segments, which can be identified in segmentation graphical

output by [] brackets around the variable name as shown in Figure 4-2. We

describe some considerations when you apply the clustering technique.

Neural clustering

As we explained earlier, the neural clustering does not require us to specify

things like similarity, but only the number of clusters we want to derive. Neural

clustering requires us to specify the maximum number of clusters. In this

particular case, the number of segments chosen are nine. The reason for

choosing nine is: the objective of behavioral segmentation is to identify the

customer characteristics and it is used as a fundamental of Customer

Relationship Management by the marketing department. Typically, dealing with

more than nine customer segments for the marketing personnel is not easy and

not efficient in terms of deploying the result.

You can specify the maximum number of segments you want to generate.

However, the decision depends on where and how the segmentation result will

be used and should be considered from a marketing perspective. For example, in

fraud detection application, we need many more segments to find abnormal

behaviors in very small segments, which could possibly exist in huge call

transaction data sets. Fraud detection in the telecoms industry will be discussed

in detail in Chapter 7, “Can you determine the characteristics of known and

unknown fraudulent behavior?” on page 131.

Note: Both algorithms generate the score (ranges from 0 to 1) which

represents how well fit each record is in the assigned cluster, in other words,

cluster membership. In demographic clustering, the record is better fitting in

the cluster if the score is closer to 1, whereas in neural clustering it is better

fitting in the cluster if the score is closer to 0.

Chapter 4. How to discover the characteristics of your customers 59

Usually neural clustering requires the input data as normalized or scaled to a

range of 0.0 to 1.0, In addition, categorical values are converted into a numeric

code for presentation to the neural network. If you have categorical fields which

has many distinct values, your performance of the mining run may be degraded

due to very large neural network.

Demographic clustering

Major difference between two clustering techniques is that the demographic

clustering technique tries to discover the number of clusters automatically, while

the neural clustering needs to specify how many clusters you want to derive. To

determine number of clusters automatically, it requires you to specify how similar

the customers within clusters should be — similarity threshold. The threshold

ranges between 0 and 1, where value of 1 means that all customers in a cluster

must be identical, and value of 0 means that customers can be completely

different. If two customers are more similar than the threshold value, then they

are candidates for being put into the same cluster.

Note: When you try to determine the optimum number of clusters, it is

important to keep in mind that the most important criteria is the availability

from business point of view.

Note to experts: The neural mining techniques in IM for Data automatically

normalizes the input data. If you are an expert on neural networks and you

want to do your own normalization, you can suppress this normalization. In

this case all input fields must have values between 0.0 and 1.0.