How to bridge spatial and temporal heterogeneity in link prediction? A contrastive method

Temporal heterogeneous networks play a vital role in capturing the dynamic evolution and inherent heterogeneity of complex real-world systems, establishing them as a pivotal research direction for link prediction. However, existing methodologies frequently fail to account for fine-grained differential distribution patterns and temporal dynamic characteristics, which are defined as spatial and temporal heterogeneity, respectively. To overcome these limitations, we propose a novel Contrastive Learning-based Link Prediction model, CLP, employing a multiview hierarchical self-supervised architecture to encode both spatial and temporal heterogeneity. Specifically, to address spatial heterogeneity, we developed a spatial feature modeling layer that captures intricate topological distribution patterns via node- and edge-level representations. Furthermore, to model temporal heterogeneity, we devised a temporal information modeling layer to characterize the evolutionary dependencies of dynamic graph topologies through time-level representations. Finally, we encoded heterogeneity in spatial and temporal distributions through contrastive learning lens, enabling a comprehensive self-supervised hierarchical relation modeling for the link prediction task. Extensive experiments on four real-world dynamic heterogeneous network datasets demonstrate that our CLP model consistently outperforms state-of-the-art baselines, achieving average improvements of 10.10%, 13.44% in AUC and AP, respectively.