TDAF: a bi-level optimization framework for CTR prediction with temporal drift adaptation

Abstract Recently, click-through rate (CTR) prediction research based on empirical risk minimization has achieved remarkable results, which assumes that training and test data follow the same distribution and optimizes CTR prediction models by minimizing prediction errors in the training data. However, this assumption does not hold in the real world. Specifically, user preferences change over time, leading to ‘temporal drift,’ where the distribution of test data differs from that of the training data. In this paper, we propose a ‘temporal drift adaptation framework’ (TDAF) for CTR prediction to cope with temporal drift. In TDAF, we devise a feature embedding predictor to learn the evolution of user preferences from historical feature embeddings and simulate feature embeddings after temporal drift. The model’s performance on simulated and real feature embeddings is improved through a novel bi-level optimization based on meta-learning, enhancing its ability to cope with temporal drift. TDAF is model-agnostic and can be widely applied to common CTR prediction models. We conduct extensive experiments, and the results show that TDAF improves the performance of CTR prediction models by an average of 0.55% across multiple datasets. Theoretical analysis and ablation studies further validate the effectiveness of TDAF. The code is available at https://github.com/lgxccc/TDAF-for-CTR-prediction/tree/main.