Augmentation-Free Self-Supervised Learning on Graphs

Inspired by the recent success of self-supervised methods applied on images, learning graph structured data has seen rapid growth especially centered augmentation-based contrastive methods. However, we argue that without carefully designed augmentation techniques, augmentations graphs may behave arbitrarily in underlying semantics can drastically change. As a consequence, performance existing is highly dependent choice scheme, i.e., hyperparameters and combinations augmentation. In this paper, propose novel augmentation-free framework for graphs, named AFGRL. Specifically, generate an alternative view discovering nodes share local structural information global with graph. Extensive experiments towards various node-level tasks, node classification, clustering, similarity search real-world datasets demonstrate superiority The source code AFGRL available at https://github.com/Namkyeong/AFGRL.