Temporal Encoding for Sequential Recommendation

In recommender systems, accurately modeling dynamic user preferences is crucial, especially for Internet enterprises with large volumes of user data. Recent sequential recommender systems based on neural networks have excelled in generating highly personalized recommendations by analyzing interaction sequences to infer user interests. However, these models often neglect temporal information, a significant part of user behavior sequences. We argue that temporal dynamics, like evolving user preferences, can potentially influence user decisions. Apart from user preference, item popularity can also contribute to a user-item interaction at a specific timestamp. To address the gap in modeling temporal information, we draw inspiration from the Fourier Series to propose a novel method for converting timestamps into temporal encodings, unveiling latent features suitable for deep networks. Our model, TE4SRec, employs this method with bidirectional attention networks to meld both sequential and temporal data from user histories. In this work, extensive experiments demonstrate that TE4SRec outperforms current state-of-the-art models across four diverse benchmark datasets, as evidenced by common evaluation metrics. An additional ablation study highlights the marked improvements attributed to our temporal encoding method. Furthermore, the visualizations of temporal approximations in TE4SRec illustrate its proficiency in effectively learning and interpreting the temporal dynamics of item popularity.