Generalize for Future: Slow and Fast Trajectory Learning for CTR Prediction

Deep neural networks (DNNs) have achieved significant advancements in click-through rate (CTR) prediction by demonstrating strong generalization on training data. However, in real-world scenarios, the assumption of independent and identically distributed (i.i.d.) conditions, which is fundamental to this problem, is often violated due to temporal distribution shifts. This violation can lead to suboptimal model performance when optimizing empirical risk without access to future data, resulting in overfitting on the training data and convergence to a single sharp minimum. To address this challenge, we propose a novel model updating framework called Slow and Fast Trajectory Learning (SFTL) network. SFTL aims to mitigate the discrepancy between past and future domains while quickly adapting to recent changes in small temporal drifts. This mechanism entails two interactions among three complementary learners: (i) the Working Learner, which updates model parameters using modern optimizers (e.g., Adam, Adagrad) and serves as the primary learner in the recommendation system, (ii) the Slow Learner, which is updated in each temporal domain by directly assigning the model weights of the working learner, and (iii) the Fast Learner, which is updated in each iteration by assigning exponentially moving average weights of the working learner. Additionally, we propose a novel rank-based trajectory loss to facilitate interaction between the working learner and trajectory learner, aiming to adapt to temporal drift and enhance performance in the current domain compared to the past. We provide theoretical understanding and conduct extensive experiments on real-world CTR prediction datasets to validate the effectiveness and efficiency of SFTL in terms of both convergence speed and model performance. The results demonstrate the superiority of SFTL over existing approaches.