ELBOTDS: Temporal Distribution Shift Generalization for Industry-Scale Recommender Systems

Abstract

Temporal distribution shift (TDS) degrades the generalization of recommender systems: models trained on past interactions face systematically different user behavior in future time windows. While incremental training on recent data is widely used in industry production, it can over-adapt to transient patterns and exacerbate instability. 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 ELBOTDS, a probabilistic framework for temporally robust recommendation. First, we identify key shifting factors on real-world production data and design a simple yet effective data augmentation strategy that simulate realistic time-varying factors to extend the training support. Second, we derive a self-supervised variational objective ELBOTDS to harness the benefits of this extended distribution while preventing representation collapse. Extensive experiments on both a public benchmark and a large-scale industrial dataset demonstrate ELBOTDS consistently improves next-period ranking performance over strong baselines, yielding a 2.33% uplift in GMV per user online and has been successfully deployed in production. Code is available at https://github.com/FuCongResearchSquad/ELBO4TDS.