IC-GAR: item co-occurrence graph augmented session-based recommendation

Session-based recommendation aims to recommend the next item of an anonymous user session. Previous models consider only the current session and learn both of the user's global and local preferences. These models fail to consider an important source of information, i.e., the co-occurrence pattern of items in different sessions. The co-occurrence patterns elicit the trajectory of other similar users and can improve the recommendation performance. We propose an Item Co-occurrence Graph Augmented Session-based Recommendation (IC-GAR) model, a novel session-based recommendation model that augments the representations of the current session with session co-occurrence patterns. IC-GAR consists of three modules: Encode Module, Session Co-occurrence Module and Prediction Module. The Encoder Module learns both of the user's global and local preference from the current session using Gate Recurrent Units (GRU). The Session Co-occurrence Module uses a modified variant of Graph Convolutional Network (GCN) to model higher order interactions between the item transition patterns in the training sessions. By aggregating the GCN representation of items of the current session, session co-occurrence representation is learned. The Prediction Module decomposes global preference, local preference and session co-occurrence to predict the probability scores of candidate items. Extensive experiments on three publicly available datasets are conducted to demonstrate the effectiveness of IC-GAR. 8.5-39.2% improvement are achieved across datasets in Precision @5, 10 and MRR@5, 10.