On the operational efficiency of different feature types for telco Churn prediction

Churn prediction in telco remains a very active research topic. Due to the uptake of social network analytics and the results of previous benchmarking studies showing a rather flat maximum performance effect of predictive modeling techniques, the focus has mainly shifted to expanding and exploring the relevant feature space. While previous studies generally agree that adding features typically increases predictive performance, they rarely discuss the accompanying issues such as data availability and computational cost. In this work, we bridge the gap between predictive performance and operational efficiency by devising a new feature type classification and a novel reusable method to determine optimal feature type combinations based on Pareto multi-criteria optimization. Our results provide several insights that can serve as a guideline for industry practitioners.