Link Prediction on Dynamic Heterogeneous Graphs via GNN with Multiple Attention Mechanisms

Abstract

Link prediction on dynamic heterogeneous graphs has attracted much attention in recent years, because it is an ideal abstract model of many problems in real application, such as the prediction of customer’s consuming behavior in electronic commerce. However, most existing link prediction methods are designed for homogeneous graphs and static heterogeneous graphs, which cannot perform well on dynamic heterogeneous graphs owing to lack of the ability of processing the dynamic information. In this paper, we propose a graph neural network (GNN) with multiple attention schemes to deal with the problem of link prediction on dynamic heterogeneous graphs. Specifically, the attention mechanism of the GNN learns the influences of the node features, the edge types and the transformation of dynamic heterogeneous graphs on link prediction. We evaluate our method on a real-world dataset which expresses the customer-commodity relations by a set of snapshots of dynamic heterogeneous graphs. Experimental result shows that our method outperforms the original method and other state-of-art baseline methods.