Graph self-supervised learning based on frequency corruption

A research team led by Haojie Li has developed a novel graph self-supervised learning method called Frequency-Corrupt Based Graph Self-Supervised Learning (FC-GSSL), accepted at The ACM Web Conference 2026. The core innovation addresses a key limitation in existing graph AI: most methods underutilize high-frequency signals and overfit to specific local patterns, which hurts generalization. FC-GSSL tackles this by strategically building corrupted graphs that are biased toward high-frequency information. It does this by corrupting nodes and edges according to their low-frequency contributions, then using these corrupted graphs as inputs to an autoencoder.

The model's supervision targets are the reconstruction of low-frequency and general features, which forces it to learn how to fuse information from multiple frequency bands effectively. The researchers further enhanced the approach by designing multiple sampling strategies to generate diverse corrupted graphs from the intersections and unions of sampling results. By aligning node representations from these different "views," the model can discover useful combinations of frequency components, reducing its reliance on any single high-frequency element and significantly improving robustness. Extensive validation on 14 benchmark datasets across three critical tasks—node classification, graph prediction, and transfer learning—demonstrated that FC-GSSL consistently outperforms existing methods, offering a more generalized and powerful framework for applications like recommendation systems and social network analysis where labeled data is scarce.