Time-Aware Squeeze-Excitation Transformer for Sequential Recommendation

Sequential recommendation models user preferences based on the sequence of user interactions. A key challenge in this context is the variability of user behavior. While Transformer-based models demonstrate unique superiority, existing models still struggle with modeling temporal information and adequately capturing users’ long-term and short-term preferences. This paper proposes a Time-Aware Squeeze-Excitation Transformer for sequential recommendation (TASESRec). The model has two salient features: (1) TASESRec preserves the continuity dependency within timestamps from both duration and spectrum perspectives using a time window function, and integrates timestamps and user interactions through a multi-layer encoder-decoder structure. (2) Given the non-uniqueness of users’ latent purchasing behavior, characterized by multiple potential purchasing behaviors, the model utilizes Squeeze-Excitation Attention (sigmoid activation) to comprehensively capture relevant items, thus enhancing prediction accuracy. Extensive experiments validate the superiority of the proposed model over various state-of-the-art models under several widely used evaluation metrics.