KGAT-SR: Knowledge-Enhanced Graph Attention Network for Session-based Recommendation
Qianqian Zhang, Zhuoming Xu, Hanlin Liu, Yan Tang · 2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI) · 2021
As a main task of session-based recommendation, next interaction (item) recommendation aims to recommend the next possible interaction (e.g., click on a song) given a session context (i.e., a list of happened interactions). This task faces the challenge of how to generate accurate recommendations when only the intra-session dependencies are available. Existing recommendation models usually fail to exploit the knowledge about items to model intra-session dependencies. Therefore, this paper addresses the problem of next item recommendation given a session context of an ordered, single-type-action, and anonymous session, and investigates how to effectively model intra-session dependencies by leveraging the knowledge from a knowledge graph (KG). We propose a novel model called Knowledge-enhanced Graph Attention Network for Session-based Recommendation (KGAT-SR). Its core idea is that the knowledge about items from a KG is exploited via knowledge graph attention network to generate a knowledge enhanced session graph (KESG). After the KESG is aggregated via weighted graph attention, the node features and graph topology in the graph are utilized to generate accurate session embedding, which can be used to recommend the next item. Experiments show that KGAT-SR significantly outperforms the state-of-the-art models for next item recommendation. KGAT-SR’s source code is available at https://github.com/hu-dske/KGAT-SR.