Relational learning for customer churn prediction: The complementarity of networked and non-networked classifiers
Authors: Wouter Verbeke, Thomas Verbraken, David Martens, and Bart Baesens
Presented at: NetMob2011: Second conference on the Analysis of Mobile Phone Datasets and Networks, Cambridge (US), 10 -11 October 2011
Keywords: Classification, customer churn prediction, social networks
This study examines the applicability of relational classification algorithms for customer churn prediction in the telco industry, and the existence and usability of non-Markovian social network effects. A range of new and adapted techniques are proposed which are designed to handle the massive size of the call graph, the time dimension, and the skewed class distribution present in a customer churn prediction setting. The proposed techniques are experimentally tested on a large-scale, real life telco data set containing both networked (call detail records data) and non-networked (customer related) information about millions of subscribers. The results indicate the existence of a limited yet highly relevant impact of social network effects on customer churn behavior, including non-Markovian network effects. A parallel setup to combine the output of a relational and non-relational churn prediction model leads to substantially improved performance and boosts the generated profits.