Exploiting Associations among Multi-Aspect Node Properties in Heterogeneous Graphs for Link Prediction

Recent years have witnessed the abundant emergence of heterogeneous graph neural networks (HGNNs) for link prediction. In heterogeneous graphs, different meta-paths connected to nodes reflect different aspects of the nodes' properties. Existing work fuses the multi-aspect properties of each node into a single vector representation, which makes them fail to capture fine-grained associations between multiple node properties. To this end, we propose a heterogeneous graph neural network with Multi-Aspect Node Association awareness, namely MANA. MANA leverages key associations among multi-aspect node properties to achieve link prediction. Specifically, to avoid the loss of effective association information for link prediction, we design a transformer-based Multi-Aspect Association Mining module to capture multi-aspect associations between nodes. Then, we introduce the Multi-Aspect Link Prediction module, empowering MANA to focus on the key associations among all, thus avoiding the negative impact of ineffective associations on the model's performance. We conduct extensive experiments on three widely used datasets from Heterogeneous Graph Benchmark (HGB). Experimental results show that our proposed method outperforms state-of-the-art baselines.