A Probabilistic Framework for Temporal Distribution Generalization in Industry-Scale Recommender Systems

2025-11-25

AI Summary

Recommendation systems suffer from temporal distribution shift (TDS), leading to long-term performance degradation. Existing invariant learning and self-supervised approaches exhibit unstable temporal generalization, representation collapse, and suboptimal data utilization. To address these issues, we propose the first probabilistic framework tailored for industrial-scale incremental learning, uniquely integrating statistical time-series analysis with causal graph modeling. We design a causal-structure-aware variational self-supervised objective—ELBO$_ ext{TDS}$—and introduce a time-varying factor resampling strategy to enhance temporal robustness. Deployed in Shopee’s product search system, our framework effectively mitigates representation degradation and significantly improves temporal generalization. Theoretical analysis establishes its consistency under TDS, while extensive experiments demonstrate superior temporal adaptability. Online A/B testing shows a 2.33% lift in average GMV per user, validating both practical efficacy and scalability.

Technology Category

• Machine Learning: Causal Learning
• Search and Optimization: Learning to Search
• Reasoning under Uncertainty: Causality

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Temporal distribution shift (TDS) erodes the long-term accuracy of recommender systems, yet industrial practice still relies on periodic incremental training, which struggles to capture both stable and transient patterns. Existing approaches such as invariant learning and self-supervised learning offer partial solutions but often suffer from unstable temporal generalization, representation collapse, or inefficient data utilization. To address these limitations, we propose ELBO$_ ext{TDS}$, a probabilistic framework that integrates seamlessly into industry-scale incremental learning pipelines. First, we identify key shifting factors through statistical analysis of real-world production data and design a simple yet effective data augmentation strategy that resamples these time-varying factors to extend the training support. Second, to harness the benefits of this extended distribution while preventing representation collapse, we model the temporal recommendation scenario using a causal graph and derive a self-supervised variational objective, ELBO$_ ext{TDS}$, grounded in the causal structure. Extensive experiments supported by both theoretical and empirical analysis demonstrate that our method achieves superior temporal generalization, yielding a 2.33% uplift in GMV per user and has been successfully deployed in Shopee Product Search. Code is available at https://github.com/FuCongResearchSquad/ELBO4TDS.

Problem

• Addresses temporal distribution shift degrading recommender system accuracy
• Overcomes unstable generalization and representation collapse in existing methods
• Enhances incremental learning for both stable and transient user patterns

Innovation

• Probabilistic framework for temporal distribution generalization
• Data augmentation resampling time-varying factors
• Self-supervised variational objective from causal structure