Knowledge-aware recommendation system based on cohesive and collaborative enhanced contrastive learning

Tianyu Cai, Siyu Sun, Anthony Tung, Yuxin Zhang, Shenggen Ju

Abstract: Introducing knowledge graphs as side information into recommendation systems has become a prominent research trend, with graph neural networks (GNNs) widely used to learn the rich information from relations in knowledge graphs. However, sparse supervised signals in recommendation systems and noise in knowledge graphs weaken the effectiveness of GNN-based models. Despite recent methods using self-supervised learning, critical challenges remain: (1) Existing methods fail to evaluate local node influence during information propagation, leading to gradual attenuation of user interest signals and additional knowledge noise in deep aggregation. (2) Exclusive focus on entity and relation information neglects the inherent local subgraph structural features of graphs, causing enhanced views to deviate from the original graph structure during encoding and thereby generating inaccurate user/item representations. To tackle these challenges, we propose a knowledge-aware recommendation system based on cohesive and collaborative enhanced contrastive learning. Specifically, we design a cohesive-collaborative mechanism to emphasize the influence of local nodes, which operates by generating two distinct graph structures: the core graph and the relevance graph. Then, a subgraph-aware GNN encoder is designed to extract subgraph structural information to learn connectivity patterns and contextual information accurately. We further conduct multi-view contrastive learning across the two constructed graphs to ensure the integration of knowledge graph information and the fusion of information between the two views. Experiments on three publicly available benchmark datasets validate the effectiveness of the proposed model.

External IDs:dblp:journals/www/CaiSTZJ26