Non-local Graph Convolutional Network

Graph convolutional network (GCN) has led to state-of-the-art performance for structured data. The superior performance would be partly due to the convolutional operations that operate over local neighborhoods. However, the distant long-range dependencies in data are still challenging to capture since it requires deep stacks of convolutional operations. Moreover, missing links in structured data might further hurt the performance. This paper introduces non-locality augmented graph convolution blocks into GCN to capture long-range or even disconnected dependencies. Specifically, we propose a dictionary-based non-locality encoding approach in which the non-local information is encoded by both graph convolution and dictionary-based implicit convolution. Unlike previous non-local approaches, our non-local block does not rely on the exhaustive computation of the relationship of data pairs. Thus, it is suitable for GCN, which typically models a large number of data samples. What's more, the proposed non-local blocks could be embedded into arbitrarily GCN architectures. We demonstrate the efficacy of our non-local block on four benchmark datasets.