Effective Knowledge-Aware Recommendation via Graph Convolutional Networks

Most existing graph neural network (GNN)-based knowledge-aware recommendation models rely on handcrafted feature engineering and do not allow for end-to-end training. As a state-of-the-art end-to-end framework, the Knowledge-aware Graph Neural Networks with Label Smoothness Regularization (KGNN-LS) model can extend GNNs architecture to knowledge graphs to simultaneously capture semantic relations between entities as well as personalized user preferences for entities/items, thereby making effective recommendation. However, we believe that KGNN-LS still has two weaknesses: (1) In KGNN-LS, the weights of the edges in the graph are determined solely by user preferences for relations without considering user’s (potential) personalized interests in entities/items. (2) The sum pooling adopted by KGNN-LS cannot effectively aggregate the most representative information of the neighborhood. In this paper, we propose the improved Knowledge-aware Graph Neural Networks with Label Smoothness Regularization (iKGNN-LS) model, which makes two improvements to KGNN-LS: (1) In iKGNN-LS, by introducing user-specific entity scoring functions, the edge weights are determined jointly by personalized user preferences for relations and for entities. (2) iKGNN-LS uses max pooling instead of sum pooling for neighborhood aggregation. Top-N recommendation experiments on three datasets show that iKGNN-LS outperforms KGNN-LS in terms of Precision@N, Recall@N, and F1-measure@N.