Dynamic heterogeneous graph contrastive learning based on multi-prior tasks

Abstract

Dynamic heterogeneous graph embedding aims to enable a variety of graph-related tasks by efficiently capturing structures and attributes over time in complex graph data. During recent decades, the self-supervised contrastive learning techniques has presented immense advantage as an approach to understanding dynamic heterogeneous graph. However, most forms of self-supervised contrastive learning approach for dynamic heterogeneous graph embedding nowadays emphasize single prior task, which results in its failure to capture multiscale knowledge in dynamic heterogeneous graph. Thus, this study presents a new self-supervised graph contrastive learning framework based on multi-prior tasks for multiscale knowledge in dynamic heterogeneous graph (MTDG). In the proposed model, this study first obtains four embedding vectors as contrastive samples using the local, global, long-term, and short-term encoders. Additionally, this study develops single-contrastive learning comprising four prior tasks to optimize the model’s ability to understand the knowledge on the local, global, long-term, and short-term of the dynamic heterogeneous graph. Besides, this study designs cross-contrastive learning comprising two prior tasks to achieve complementary knowledge between the four embedding vectors generated by the encoders. Furthermore, this study introduces slight random noise and a shuffling strategy to prevent the generation of similar contrastive samples. Ultimately, this study assesses MTDG on twelve dynamic graph datasets from the real world, and the experimental results show that the proposed model achieves the better results on all datasets compared to the eleven baselines.