Dynamic time-aware collaborative sequential recommendation with attention-based network

Abstract

A natural way of user modeling in sequential recommendation is to capture long-term and short-term preferences, respectively, given user historical behaviors and then fuse them together. Most existing approaches building on attention-based network focus only on exploring item–item relations within each user sequence and ignore collaborative relations among different user sequences, which restricts the improvement of recommendation quality, especially on sparse datasets. Moreover, construction and utilization of collaborative signals including the integration with the original information greatly impact the recommendation effects. In this paper, we propose a novel method named dynamic time-aware collaborative sequential recommendation with attention-based network(DTCoSR) to further address the issues. Specifically, we first design a time-aware collaborative item module to gain collaborative item representations for both long- and short-term interests, consisting of neighborhood selection and neighborhood information aggregation. Then, we utilize two independent self-attention networks to extract the two different levels of short-term interests dependent on the item representation and collaborative item representation, respectively, and then adaptively merge them as the final short-term interests. We achieve long-term interest via the correlation between the user embedding and its collaborative item embedding. Finally, DTCoSR fuses long- and short-term interests in an adaptive method. Extensive experiments on three real-world datasets show that DTCoSR outperforms state-of-the-art methods.