A0441
Title: Using feature selection based on multivariate statistical dependence for churn prediction in the automotive industry
Authors: Thimo Kasper - University of Salzburg (Austria) [presenting]
Laura Koenig - Porsche Informatik GmbH (Austria)
Markus Gruber - Porsche Informatik GmbH (Austria)
Thomas Soboll - Porsche Informatik GmbH (Austria)
Wolfgang Trutschnig - University of Salzburg (Austria)
Abstract: In recent years customer retention and preventing customer churn is of increasing importance for businesses in various industries. Particularly in the non-contractual automotive aftersales market, where churn is a not directly observable latent event, retaining customers prone to defecting is key for the profitability of dealerships and workshops. Therefore, working with real-life data from a group of workshops/dealerships from an international automotive distributor in an Austrian region, we tackle the question of how to assess aftersale customer churn probabilities by using techniques from statistics and machine learning in order to allow for dynamic customer selection and targeted retention marketing. Driven by the need for efficient, well interpretable models, special focus is assigned to the feature selection procedure - four recently developed methods based on bivariate and multivariate statistical dependence are benchmarked against random forest's feature importance and a selection based upon Pearson's correlation coefficient. Our findings show that reliable, well-performing customer churn prediction with low model complexity is indeed possible in the context of automotive aftersales.
