Building comprehensible customer churn prediction models with advanced rule induction techniques

  • Authors:

  • Wouter Verbeke;David Martens;Christophe Mues;Bart Baesens

  • Affiliations:

  • Department of Decision Sciences and Information Management, Katholieke Universiteit Leuven, Naamsestraat 69, B-3000 Leuven, Belgium;Department of Decision Sciences and Information Management, Katholieke Universiteit Leuven, Naamsestraat 69, B-3000 Leuven, Belgium and Department of Business Administration and Public Management, ...;School of Management, University of Southampton, Highfield Southampton, SO17 1BJ, United Kingdom;Department of Decision Sciences and Information Management, Katholieke Universiteit Leuven, Naamsestraat 69, B-3000 Leuven, Belgium and School of Management, University of Southampton, Highfield S ...

  • Venue:
  • Expert Systems with Applications: An International Journal
  • Year:
  • 2011

Quantified Score

Hi-index 12.06

Visualization

Abstract

Customer churn prediction models aim to detect customers with a high propensity to attrite. Predictive accuracy, comprehensibility, and justifiability are three key aspects of a churn prediction model. An accurate model permits to correctly target future churners in a retention marketing campaign, while a comprehensible and intuitive rule-set allows to identify the main drivers for customers to churn, and to develop an effective retention strategy in accordance with domain knowledge. This paper provides an extended overview of the literature on the use of data mining in customer churn prediction modeling. It is shown that only limited attention has been paid to the comprehensibility and the intuitiveness of churn prediction models. Therefore, two novel data mining techniques are applied to churn prediction modeling, and benchmarked to traditional rule induction techniques such as C4.5 and RIPPER. Both AntMiner+ and ALBA are shown to induce accurate as well as comprehensible classification rule-sets. AntMiner+ is a high performing data mining technique based on the principles of Ant Colony Optimization that allows to include domain knowledge by imposing monotonicity constraints on the final rule-set. ALBA on the other hand combines the high predictive accuracy of a non-linear support vector machine model with the comprehensibility of the rule-set format. The results of the benchmarking experiments show that ALBA improves learning of classification techniques, resulting in comprehensible models with increased performance. AntMiner+ results in accurate, comprehensible, but most importantly justifiable models, unlike the other modeling techniques included in this study.