HIDyG: heterogeneous information perception in dynamic graphs via contrastive learning

Runshuo Liu, Chao Li, Qianyu Song, Hui Zhou, Qingtian Zeng

Abstract: Dynamic graphs can accurately portray the evolution of network structures over time in realistic scenarios. Previous continuous-time dynamic graph approaches focus on processing time-series information, ignoring the heterogeneous semantic information of nodes. In addition they fail to capture the frequent dynamic changes among higher-order neighbors. To solve above limitations, in this paper, a Heterogeneous Information Perception in Dynamic Graphs via Contrastive Learning (HIDyG) method is proposed. We propose a heterogeneous information perception strategy. It utilizes meta-paths to capture heterogeneous semantic information of dynamic nodes from different perspectives. By combining temporal analysis and structural diffusion strategy, it can effectively identify frequent higher-order changes during network evolution. Additionally, a dynamic graph collaborative contrastive learning method is designed. It leverages meta-paths to construct positive samples to guide the embedding learning process between source nodes and target nodes. By iteratively refining the learned node representations, it ensures effective adaptation to downstream tasks. Empirical evaluations across five benchmark datasets substantiate the superior performance of the proposed HIDyG framework in addressing both link prediction and node classification tasks. Specifically, HIDyG achieves significant improvements in transductive link prediction with AP scores of 99.10% on Reddit, 99.05% on Wikipedia. For inductive learning, HIDyG attains AP scores of 98.76% on Reddit and 98.80% on Wikipedia. The source code of this work is available at https://github.com/LiuRS1/HIDyG.

External IDs:dblp:journals/apin/LiuLSZZ25