Customer churn is considered a critical issue for all businesses, as customer loss leads to a decrease in future profits. Although acquiring new customers can address such losses, this process tends to cost more than retaining existing customers. Therefore, identifying potential churners and then retaining them is important. While churn prediction has been studied widely, current research must analyze more complex business models in response to their increase, such as platform businesses. However, modeling churn prediction for these businesses is challenging because consumer behavior over platforms is more complicated. Past approaches to churn prediction can be improved upon using recent advancements in deep learning that capture the nonlinear relationships behind the data in a data-driven manner, especially for complex business models. This study proposes a method of extracting latent features from purchase histories as explanatory variables for churn prediction using a variational autoencoder with the actual customer distribution as a prior. The proposed method is validated using real purchase data from a platform business and shows a 1.5% improvement in F-measure against the baseline, and a 20% improvement for customers with recent transactions. Subsequently, the variables are examined using several methods for data analysis to interpret the meanings of the extracted features and underlying customers’ purchasing behavior in the group of potential churners. With these analyses, the model provides practical implications for understanding what kind of purchasing behavior may lead to churn and planning effective retention strategies.
ASJC Scopus subject areas