Bridging Technical and Business Evaluation in E-Commerce Churn Prediction: The PDCL Labeling and Hybrid eCCP-Net Approach

In the digital era, business decisions are increasingly guided by data-driven approaches, transforming information into insights and actions aimed at enhancing revenue and customer lifetime value (CLV) for sustainable growth. To achieve long-term customer retention, companies must deeply understand customer behavior and predict future actions, as acquiring new customers is considerably more expensive than retaining existing ones. However, both public and enterprise datasets often lack explicit churn labels, while many previous studies depend on overly simplistic labeling methods that fail to reflect actual customer behavior, leading to high-performing models with limited business relevance. This paper introduces a novel Personalized Dynamic Churn Labeling (PDCL) method that captures individual purchasing behaviors and product category characteristics to generate precise churn labels, and proposes Hybrid eCCP-Net (e-commerce Customer Churn Prediction Network), a multi-modal deep learning model capable of integrating static, sequential, and textual customer data for deeper behavioral insights. Beyond standard technical metrics, this study emphasizes a business-oriented evaluation metric as the primary assessment criterion to align predictive modeling with real-world financial outcomes. Experimental results on the Brazilian E-Commerce Public Dataset by Olist demonstrate that PDCL significantly improves upon existing pre-labeling methods, while Hybrid eCCP-Net achieves the best business-oriented performance, confirming its effectiveness in improving customer retention and reducing misclassification costs in practical e-commerce applications.