CPGCN: Collaborative Property-aware Graph Convolutional Networks for Service Recommendation

Collaborative filtering (CF) recommendation methods usually suffer from data sparsity and the cold start problem. An effective way to deal with these problems is to add side information such as knowledge graph (KG). However, the existing knowledge-aware recommendation methods cannot separately and accurately model users' interest preferences and services' property characteristics, uniformly termed as property features in this paper. They also rely too much on semantic information of KG and ignores the collaborative information between users and services which plays an important role in CF-based methods. To address these issues, we propose an end-to-end model called Collaborative Property-aware Graph Convolutional Networks (CPGCN) to learn the fine-grained property features. Firstly, the user-service collaborative information is fused with the semantic information of KG to construct Collaborative Property-aware Graphs (CPG). Then property features are learned by one kind of effective variant of graph convolutional neural network (GCN) in recommendation scenario. In order to reflect the differences among users and services, a relevance-attention network is proposed to aggregate these property features. Through extensive experiments on real-world service scenarios, we have confirmed the superiority of CPGCN compared with other state-of-the-art models. Each module in CPGCN is also confirmed to play an indispensable role in recommendation.