Towards Better Dynamic Graph Learning: New Architecture and Unified Library

We propose DyGFormer, a new Transformer-based architecture for dynamic graph learning that solely learns from the sequences of nodes' historical first-hop interactions. DyGFormer incorporates two distinct designs: a neighbor co-occurrence encoding scheme that explores the correlations of the source node and destination node based on their sequences; a patching technique that divides each sequence into multiple patches and feeds them to Transformer, allowing the model to effectively and efficiently benefit from longer histories. We also introduce DyGLib, a unified library with standard training pipelines, extensible coding interfaces, and comprehensive evaluating protocols to promote reproducible, scalable, and credible dynamic graph learning research. By performing extensive experiments on thirteen datasets from various domains for transductive/inductive dynamic link prediction and dynamic node classification tasks, we observe that: DyGFormer achieves state-of-the-art performance on most of the datasets, demonstrating the effectiveness of capturing nodes' correlations and long-term temporal dependencies; the results of baselines vary across different datasets and some findings are inconsistent with previous reports, which may be caused by their diverse pipelines and problematic implementations. We hope our work can provide new insights and facilitate the development of the dynamic graph learning field. All the resources including datasets, data loaders, algorithms, and executing scripts are publicly available at https://github.com/yule-BUAA/DyGLib.