Context-aware item attraction model for session-based recommendation

Session-based recommendation uses existing items in users’ interaction sessions to predict the next items with which users will interact. The existing items in sessions usually have different degrees of relevance with each other, and this item relevance also reflects users’ interests. Moreover, when sessions are represented in different structural forms, there will be different types of relevance between items, an aspect typically neglected by previous work. In this paper, we propose a novel Context-aware Item Attraction Model (CIAM) for session-based recommendation, which is capable of capturing different types of relevance between items in order to obtain users’ general and temporal interests and predict the next items in sessions. First, we convert sessions into local and global undirected graphs to mine the item adjacency relevance within and across sessions in order to better determine users’ general interests. Second, we retain the natural sequence structure of sessions, and model the transition relevance between items in sessions to get users’ temporal interests. Third, we design a context-aware item embedding method to obtain the embedding of each item; this method utilizes superposition and a weighted graph convolutional network to aggregate the context information from both the item's features and the item's neighborhood. Finally, based on users’ general and temporal interests, as well as the context-aware embeddings of items, we predict the next items with which users will interact during a session. The proposed model is then extensively evaluated on two real-world datasets. Experimental results show that our model outperforms the state-of-the-art baseline methods. Through the analysis of the experiments, we prove that our model can effectively capture the different types of relevance between items within and across sessions for accurately modeling user interests, therefore improving recommendation performance.