Knowledge-aware Graph Convolutional Network for Collaborative Recommendation

Abstract

Recommender systems(RS) that introduce the knowledge graph(KG) overcome cold-start and data sparsity problems through mining the abundant side information. Recent popular KG-based recommendation methods are mainly propagation-based methods, however, such methods often fail to fully exploit the rich semantic associations between entities and ignore the potential key signals in user' actions. Therefore, we use a knowledge-aware graph convolutional network (GCN) for collaborative recommendation (KACR), which first explicitly encodes collaborative signals by collaboratively propagating user-item interaction data. Secondly, we use a graph convolutional network to aggregate semantic information of the local neighborhood structure and propagate users' interest in the KG's high-order connected structure to mine the users' potential preference. Finally, users' and items' final representation is obtained through the aggregator to make predictions. According to the experimental findings on three open datasets, KACR has a stronger recommendation effect than the baselines.