Churn Prediction in Gasoline Consumers under the Price Commitment Scenario

Boyang Li, Yunzhe Qiu, Lili Chen, Xi Zhang

TL;DR: This paper introduces the EFANPC model, which uses multi-source domain adaptation to predict customer churn under price commitments in the gasoline market, even without historical data in the scenario.

Abstract: The gasoline consumption market demonstrates significant brand loyalty, with consumers often preferring specific gas stations. To enhance customer retention and manage price fluctuation impacts, some companies have adopted price commitment strategies, assuring consumers they will not pay more than a predetermined future price, thus mitigating purchasing risks. However, predicting customer churn in the price commitment scenario is challenging due to the lack of historical churn data specific to these scenarios. To address the problem, we introduce the Enhanced Feature Adaptation Network in Price Commitments (EFANPC) model. The EFANPC model employs multi-source domain adaptation (MDA) techniques to transfer knowledge from various source domains without price commitments to the target domain with price commitments. It incorporates a newly designed loss function that considers domain distances in both price commitment and regular scenarios, effectively addressing the unsupervised customer churn prediction problem under price commitments. We develop features that reflect consumer purchasing behaviors and introduce a feature selection method combining both common and domain-specific features. This method captures the unique consumer behavior characteristics related to churn under price commitments for each source domain. Additionally, to tackle the challenges of insufficient samples and class imbalance in the source domain, we propose a method that balances class weights and utilizes samples from all source domains for each classifier's learning, enhancing predictive performance. The EFANPC model's performance is validated through a case study in North China, demonstrating its effectiveness in predicting churn and offering practical insights for gasoline companies.

Submission Number: 149