Personalized Quality-Aware Reranking for Long-Term User Retention in E-commerce

Abstract

Modern recommendation systems are predominantly optimized for short-term engagement signals such as Click-Through Rate (CTR) and Conversion Rate (CVR), often overlooking post-purchase user satisfaction. As a result, ranking systems can repeatedly promote products that drive immediate interactions but lead to poor user experiences and long-term dissatisfaction. We address this gap through a personalized quality-aware reranking framework that explicitly incorporates product quality into the ranking objective. We adopt the Net Quality Detractor (NQD) ratio, the fraction of 1–2 star post-purchase ratings, as a direct measure of severe dissatisfaction that average rating alone cannot capture. Building on this signal, we propose a dual-model architecture: Personalized NQD (PNQD), which estimates user-specific dissatisfaction risk for a user–product pair, and Retention-Driven Alpha (RDA), which adaptively controls the strength of the quality penalty based on a user's historical experience and predicted retention. Offline experiments show that PNQD captures user-level quality preferences beyond product-level signals, while large-scale online A/B experiments demonstrate a 0.48% reduction in realized post-purchase NQD and 0.47% Net Merchandise Value (NMV) uplift over a non-personalized quality baseline, with neutral impact on engagement. The system is fully deployed, serving approximately 200 million monthly active users.