Dynamic Sequential Recommendation: Decoupling User Intent from Temporal Context

Abstract

Nowadays sequential recommender systems have been equipped with various deep learning techniques such as recurrent neural networks and self-attention mechanisms. Despite recent advances, the most striking difference between the field of recommendation and other fields of study—the dynamic property—has often been ignored when modeling user behavior. In contrast to the case of language models where word frequency is relatively stable, the item popularity distribution changes periodically or irregularly. To address this time-related problem, we propose a novel time-aware framework for dynamic sequential recommendation. More precisely, based on factorizing the point-wise mutual information between the item and the context, the user intrinsic interest is decoupled from the temporal context. The sequential model consists mainly of two parts: i) a time-invariant main network, intricately structured with most of the model parameters and stable to temporal dynamics; ii) a time-sensitive bias network, frequently updated to represent the evolution of item popularity. The proposed strategy can be overlaid on the state-of-the-art sequential models to simultaneously capture the sequential and temporal patterns. Furthermore, the connection with the matrix factorization is illustrated to intuitively demonstrate the decoupling of user intent. Finally, the effectiveness of the proposed strategy on retrieval and ranking metrics is demonstrated through extensive experiments from several real-world scenarios, including MovieLens, Amazon and Alibaba datasets.