Temporal Edge-Aware Hypergraph Convolutional Network for Dynamic Graph Embedding

Abstract

Graph embedding is a critical aspect of network analysis that helps to advance various real-world applications such as social recommendation and protein structure prediction. Most of the existing graph embedding methods are designed for static graphs while many real-world graphs intrinsically behave as dynamic graphs. Recent works try to combine graph neural networks(GNN) with recurrent neural networks to address this issue. However, these methods can not independently utilize GNN models to cope with dynamic graphs and they ignore the inner edge-level correlations in dynamic graphs. To tackle these problems, we propose a novel dynamic graph embedding framework in this paper, called DynHyper. Specifically, we introduce a temporal hypergraph construction to capture the local structure information and temporal dynamics simultaneously. Then, we employ a hyperedge projection to obtain edge-level correlations. Further, we propose a temporal edge-aware hypergraph convolution to transmit and aggregate the messages in the temporal hypergraph. We conduct our experiments on seven real-world datasets to evaluate the effectiveness of DynHyper in both link prediction and node classification tasks. Experimental results show that DynHyper significantly outperforms all baselines, especially on the more complex datasets.