Attenuated sentiment-aware sequential recommendation

Sequential recommendation(SR) focuses on modeling the historical relationship of a user’s behavior. The attention-based models such as Transformer and BERT have been introduced in SR and acquired excellent performance. However, these models mostly only utilize the user-item interaction sequential data but ignore the additional information. We argue that the complicated human subjective sentiment plays an essential influence on their consuming behavior. In this paper, we introduced attenuated sentiment information into sequential recommender to capture user potential preference. Specially, we propose an attenuated sentiment memory network (ASM-Net) to simulate the real decay of human sentiment according to the time interval relationship. We construct a two channels recommender architecture called attenuated sentiment sequential recommendation (ASSR) to generate user sentiment preference and item preference. Specifically, the first channel models the general item attention-aware sequential relationship and the secondary channel utilizes multi-attenuated sentiment-aware attention to capture sequential preference. We collect two industrial Chinese datasets and two open English datasets to verify the model’s performance. We design ablation study and sentiment sensitivity to investigate the influence of attenuated sentiment on user preference. Comprehensive experimental results demonstrate that our sentiment decay modeling approach is effective to capture users’ subjective preferences, and our method outperforms several state-of-the-art recommenders.