SEGAR: Knowledge Graph Augmented Session-Based Recommendation

Abstract

Predicting the next interaction item in the session-based recommendation system is an emerging and challenging research task. Existing studies model a session as a sequence or graph of items for predicting the next-click item. However, these approaches ignore the global graph-based relations between the session items and the local neighborhood-based item relevance to external knowledge bases, thus fail to encode rich semantic knowledge between items for achieving comprehensive and accurate recommendations. To overcome the current shortcomings, we proposed a novel knowledge graph augmented model called SEGAR (Knowledge Graph Augmented Session-based Recommendation) by leveraging graph convolutional network and knowledge graph attention network. When integrating the static local attributes and the knowledge about all the last session items encoded in their k-hop neighborhoods in the knowledge graph, SEGAR models all sessions as a session graph and captures the dynamic global temporal and popularity-aware information from the session context. The model encodes a comprehensive semantic knowledge between items for achieving more accurate recommendation. Extensive experiments on two benchmark datasets show that SEGAR outperforms four state-of-the-art models on the session-based recommendation task.