DynGCN: A Dynamic Graph Convolutional Network Based on Spatial-Temporal Modeling

Abstract

Representation learning on graphs has recently attracted a lot of interest with graph convolutional networks (GCN) achieving state-of-the-art performance in many graph mining tasks. However, most of existing methods mainly focus on static graphs while ignoring the fact that real-world graphs may be dynamic in nature. Although a few recent studies have gone a step further to incorporate sequence modeling (e.g., RNN) with the GCN framework, they fail to capture the dynamism of graph structural (i.e., spatial) information over time. In this paper, we propose a Dynamic Graph Convolutional Network (DynGCN) that performs spatial and temporal convolutions in an interleaving manner along with a model adapting mechanism that updates model parameters to adapt to new graph snapshots. The model is able to extract both structural dynamism and temporal dynamism on dynamic graphs. We conduct extensive experiments on several real-world datasets for link prediction and edge classification tasks. Results show that DynGCN outperforms state-of-the-art methods.