Towards better recommendations: Integrating counterfactual learning and trust regions in digital platforms

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. • Added GRU4Rec Benchmark Comparison: Tables 6 and 7 now include f ̂ GRU 4 Rec and f ̂ random baseline comparisons, demonstrating that our proposed method substantially outperforms established session-based recommendation approaches across all session sizes (2–6). • Cross-Model Robustness Validation: New analysis examining error propagation between Logistic Regression and LightGBM models shows that sessions optimized using LightGBM maintain robust predictions when independently re-scored, validating framework reliability against model uncertainty. • Enhanced Key Concept Definitions: Added explicit definitions for critical terms, including implicit feedback (“user behaviors that indirectly signal preferences such as dwell time”), validated domain, trust region regularization, and session (distinguished from HTML web sessions). • Strengthened Literature Foundation: Added matrix factorization citations and moved OPPM explanation earlier with concrete examples, providing clearer theoretical grounding before discussing limitations. • Streamlined Content and Practical Implications: Removed tangential material on adversarial behavior; added statement about framework’s “plug and play” flexibility for different predictive models and noted feasible extensions including error-aware optimization and decision-focused learning.