Towards better recommendations: Integrating counterfactual learning and trust regions in digital platforms
提出一个两阶段框架，利用反事实学习和信任区域约束直接优化会话级点击率，解决推荐系统目标与平台业务指标错位的问题，在网易云音乐数据上取得显著提升。
Most recommender systems optimize individual item preferences rather than session-level business metrics, misaligning algorithmic objectives with platform goals. We propose a two-stage framework that directly optimizes session-level click-through rates (CTR) using counterfactual learning and trust-region constraints. Stage one trains models to predict positive session outcomes using collaborative filtering features. Stage two optimizes over these models with trust-region regularization to find alternative sessions that maximize expected CTR while ensuring prediction reliability. Using NetEase Cloud Music sessions, our LightGBM-based framework delivers substantial CTR gains across session sizes while staying within validated domains. It enables direct session-level optimization, integrates robust feedback, and applies trust regions, providing practical, business-aligned recommendations.