SR-DSGA: Session Recommendation for Dual Sequence Based on Graph Neural Network and Multi-Attention

Abstract

Session recommender system (SRS) captures user’s sequential features based on historical behavior to predict the next-clicked item. The accuracy of extracting user’s session features directly determines the key performance of SRS. Existing session recommendation methods have two flaws: 1) ignore the complex connections between items, i.e. represent them in a relatively isolated manner; 2) neglect the transition patterns between attributes of items. To address these issues, we propose a novel session recommendation model named SR-DSGA (Session Recommendation for Dual Sequence based on Graph neural network and multi-attention). Firstly, SR-DSGA adopts message passing mechanism in graph neural network to get non-isolated item embedding representations with specific semantic relationship by item-level explicit sequence modeling. Secondly, SR-DSGA exploits the Transformer’s multi-head self-attention mechanism to indirectly obtain item embedding representations in another way through item attribute-level implicit sequence modeling. Therefore, SR-DSGA can help extract the fine-grained features with full sequential patterns even in sparse data scenarios. Finally, soft-attention and time threshold are used to acquire user’s long-term and short-term preferences respectively. Experimental studies on real-world datasets demonstrate the proposed SR-DSGA model outperforms the state-of-the-art benchmark methods.

Keywords

• Session recommendation
• behavior features
• dual sequence modeling
• graph neural network
• multi-attention