Churn Prediction in Telecommunications Using MiningMart

This paper summarises a successful application of Knowledge Discovery in Databases (KDD) in an Italian telecommunications research lab. The aim of the application was to predict customer churn behaviour. A critical success factor for this application was clever preprocessing of the given data, in particular the construction of derived predictor features. The application was realised in the MiningMart KDD system, whose particular strength is data preprocessing on a conceptual level. Since MiningMart provides a declarative, yet executable model of the presented application, this model could be published in a central repository of KDD models, where it is publicly inspectable, which complements the descriptions in this paper. 1 A case study in churn prediction A major concern in customer relationship management in telecommunications companies is the ease with which customers can move to a competitor, a process called “churning”. Churning is a costly process for the company, as it is much cheaper to retain a customer than to acquire a new one [1]. The objectives of the application to be presented here were to find out which types of customers of a telecommunications company are likely to churn, and when. The task was solved using decision trees which achieved a predictive accuracy of 82%. This good result was only possible due to the introduction of relevant derived features for prediction which were not available in the original data, and due to a rerepresentation of the data so that temporal aspects could be included. Thus data preprocessing was a key success factor in this application. During preprocessing, the available data tables were transformed so that a classification algorithm could be applied. In the resulting data set, each row (that is, each example for classification) corresponded to one customer of the company, and contained many features describing their telecommunication behaviour for each of five consecutive months. Whether or not the customer left the company in the sixth month determined the binary classification label or target. Once a learned classifier is available it can be applied every month on data from the current and past four months, to predict churn for the following month.