CARE: An Integrated Framework for Contagion-Aware Churn Prediction and Network-Level Risk Quantification

Le résumé fourni par la source

Customer churn costs digital platforms billions annually, and growing evidence shows that churn exhibits contagion-like patterns across social networks, where a single departure is associated with elevated churn risk among connected users. However, existing prediction methods either model users independently or aggregate neighbor features without distinguishing churned from active connections. This leaves two coupled problems unresolved: the absence of contagion-aware prediction and the lack of network-level risk quantification. To address these problems, we propose CARE (Contagion-Aware Retention Ecosystem), a unified framework comprising two modules. The Churn-aware graph convolutional network introduces a learnable contagion modulation factor into graph attention that selectively amplifies messages from churned neighbors that are predictively informative of subsequent churn. The Churn Contagion Index, derived from first-order spillover decomposition, integrates churn probability, social influence centrality, and lifetime value to distinguish high-impact super-spreaders from isolated churners. Together, the two modules bridge the gap between individual churn prediction and network-level retention optimization. Experiments on the KuaiRec dataset show that the proposed framework achieves an area under the curve of 0.921, outperforming ten baselines by +2.9 percentage points, and that index-based targeting achieves 44.9% higher second-order churn rate capture on a held-out temporal window, consistent with retained high-index users co-occurring with reduced downstream churn beyond the training period. Further evaluation on the KKBOX dataset demonstrates generalization to behavior-inferred graphs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.