Subgraph Convolutional Network for Recommendation

Nowadays recommendation systems play an important role in our lives, which help users to quickly identify their desirable items. The networking trend for world (i.e., every sector of our world can be networked) has made the recommendation systems one of the intensively studied research areas in the last decades. In this paper, we formulate a graph-based recommendation problem, which aims to find the most relevant nodes for a given set of query nodes in the graph. For graph embedding, Graph Convolutional Network (GCN), which aggregates neighbor information via convolution layers, is an effective model. However, a convolution layer in a GCN only considers unstructured information, i.e., it takes single nodes as input and only leverages the first-order neighbor information, so only limited local information can be learned. To overcome the mentioned limitations, we develop a Subgraph Convolutional Network (SCN) model which aggregates both neighbor node information and structural information via convolution layers. Moreover, the fully connected layer based link prediction is integrated for effective recommendations. The experimental results on real-world datasets verify the effectiveness of our solution.