Churn comprehension analysis for telecommunication industry using ALBA

Customer churn analysis is getting immense attention from the business community, especially in the telecommunication sector, to help in generating more revenue. This raises the need for modeling a churn prediction system that is not only accurate but also encompasses comprehensibility and justifiability. Also the judgment ability of the domain expert to identify the drivers of churn and determine a correlation between those drivers plays a very important role. For the stakeholder or businessman, it is very important to predict the main drivers of churn to increase the profit. In this paper, we use a technique called Active Learning Based Approach (ALBA), with two rule induction methods, Decision Tree and Ripper to develop a churn prediction model that is not only accurate but also yields comprehensibility and justifiability. ALBA is a rule extraction technique based on a flavor of Support Vector Machine (SVM) called Active Learning. Basically, this approach is used as instance generator. Using this approach, we first remove noise from the data by replacing the SVM predicted labels with the original labels and secondly, we generate artificial instances in the noisy area near support vectors. These newly generated instances are not class-wise balanced. For this purpose, we use SMOTE to make them balanced.