GSTBench: A Benchmark Study on the Transferability of Graph Self-Supervised Learning Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Inform
Yu Song, Zhigang Hua, Yan Xie, Jingzhe Liu, Bo Long, Hui Liu · 2025
Read GSTBench: A Benchmark Study on the Transferability of Graph Self-Supervised Learning Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Inform by Yu Song, Zhigang Hua, Yan Xie, Jingzhe Liu, Bo Long, Hui Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-06975.txt.
Description
Graph learning methods that teach themselves from unlabeled data mostly fail to transfer to new datasets, finds a 2025 benchmark study led by Yu Song, with five co-authors. Self-supervised graph methods are usually developed and tested on one dataset at a time. Their cross-dataset transfer, the skill foundation models need, was largely unexplored.
GSTBench is presented as the first systematic benchmark for that transfer. Five representative methods were pretrained at large scale on a hundred-million-paper citation graph, then evaluated across diverse target graphs under a standardized setup that isolates the pretraining objective. Surprisingly, most methods generalized poorly, some doing worse than random initialization. One masked autoencoder approach, GraphMAE, improved transfer consistently. The authors analyze what drives the difference and offer guidance for building graph methods on a pretrain-then-transfer model.
Catalog record
| CREATOR | Yu Song, Zhigang Hua, Yan Xie, Jingzhe Liu, Bo Long, Hui Liu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-06975