Heterogeneous Temporal Graph Neural Networks for Link Prediction

Link prediction stands as a critical research frontier in graph data analytics. In real-world scenarios, graph data commonly manifest heterogeneity and temporal dynamism, thereby adding layers of complexity to analytical models. Existing work either considers heterogeneity or temporal dynamism, with very little work modeling both properties simultaneously. In this study, we propose a novel link prediction framework that is specifically designed to integrate heterogeneity and temporal variations within graph data, which can be used in a wide range of domains. In the framework, we design a Heterogeneous Temporal Graph Neural Network (HTGN) model and a new training strategy, named Bi-Window Strategy(BWS). The HTGN is capable of synthesizing temporal node representations by aggregating heterogeneous temporal information. Simultaneously, the Bi-Window Strategy (BWS) enhances the model's ability to capture the long-term distributional characteristics inherent in graph data. Our approach achieved 2nd-Winner Award of Temporal Link Prediction task at WSDM'22 Cup. In this paper, we incorporate additional experiments using a more equitable data-splitting approach and perform comparisons with an expanded range of baseline methods.