Feature-enhanced embedding learning for heterogeneous collaborative filtering

Heterogeneous information network (HIN) has recently been receiving increasing attention in recommender systems due to its practicability in depicting data heterogeneity. The rich structural and semantic information embodied in the HIN can help mining latent features of users and items for recommendations. However, almost all existing HIN-based recommendation methods focus on the design of complicated learning architecture while using simply initialized features. In this paper, we propose a novel feature-enhanced embedding learning model which combines informative feature initialization strategy with simple learning architecture for heterogeneous collaborative filtering. We first build multiple homogeneous sub-networks by extracting different relations guided by meta-paths from the HIN. We then design a comprehensive feature initialization strategy that contains semantic and spatial encoding module to characterize the node feature. After that, a simple learning architecture based on multi-layer perceptron is applied to learn the latent representation of users and items. Next, a novel convolutional neural network-based fusion mechanism is used to determine the attention weight of semantic relations and compress multiple embedding vectors into a compact representation to apply for final recommendation. Finally, we conduct extensive experiments on two classic datasets to demonstrate the effectiveness and feasibility of the proposed FHetCF method in solving HIN-based recommendation tasks. Results show that the proposed method soundly outperforms the competitive baselines by 1.71 to 10.46% on hit ratio and 3.17 to 13.75% on normalized discounted cumulative gain, respectively. The proposed method opens up a new avenue to effectively utilize heterogeneous information to improve recommendation performance.