A1004
Title: Predictive subgroup logistic regression: A new approach in customer churn modeling with unobserved heterogeneity
Authors: Rui Huang - Nanjing University (China) [presenting]
Kun Chen - Southwestern University of Finance and Economics (China)
Zhiwei Tong - The University of Iowa (United States)
Abstract: Modeling customer churn has become increasingly vital in the competitive landscape of today's markets, as it helps businesses understand customer churn behavior and develop tailored marketing strategies. Traditional techniques, such as decision trees and logistic regression, often overlook the heterogeneity in the latent factors driving customer churn. The aim is to introduce a novel predictive subgroup logistic regression (PSLR) model designed to identify unobserved subgroup structures among existing customers, accurately classify new customers into these subgroups and subsequently generate churn predictions. A penalized likelihood function is derived and optimized for estimating this model, addressing the challenges associated with optimization through the development of an alternating direction method of multipliers (ADMM) algorithm, for which convergence is proven. Extensive simulation studies confirm the PSLR model's efficacy in inferential and predictive tasks. An empirical study of a telecommunication dataset demonstrates that the PSLR model identifies the presence of unobserved heterogeneity among customers, even after initial segmentation by decision trees. Moreover, compared to selected benchmark models, the PSLR model achieves more balanced performance in identifying both positives and negatives while also achieving better or at least comparable results in terms of various aggregate accuracy metrics.
