Author:

Abstract:

In many applications, identifying potential churners is of great importance and has been widely studied. Recently, literature has acknowledged the power of social network analysis for churn detection, which has been proven to achieve more accurate results. We focus on churn detection in the telecommunications industry, where a social network is constructed based on call records. We contribute by evaluating and comparing two community detection approaches and, as a result, identify the effects of peer pressure on the likelihood of individuals to churn. Particularly, we propose a two-step procedure. In a first step, we detect the relevant communities of the social network using two different methods: (1) a top-down clustering approach, and (2) a bottom-up clustering approach. The top-down clustering approach results in few and large clusters, whereas the bottom-up clustering identifies complete cliques and hence produces smaller but a greater number of clusters. In a second step, we enrich churn prediction models, which traditionally only use intrinsic features. From the clusters, we extract community features and use them as additional variables to predict customer churn. Finally, we benchmark both above mentioned community detection approaches to results from the whole, un clustered network and determine which of the clustering techniques excels. Our results show how pre-clustering techniques boost the performance of churn prediction methods.