Contrastive Self-supervised Learning for Graph Classification

Graph classification is a widely studied problem and has broad applications. In many real-world problems, the number of labeled graphs available for training models limited, which renders these prone to overfitting. To address this problem, we propose two approaches based on contrastive self-supervised learning (CSSL) alleviate first approach, use CSSL pretrain graph encoders widely-available unlabeled without relying human-provided labels, then finetune pretrained graphs. second develop regularizer CSSL, solve supervised task unsupervised simultaneously. perform graphs, given collection original data augmentation create augmented out An created by consecutively applying sequence alteration operations. A loss defined learn judging whether are from same graph. Experiments various datasets demonstrate effectiveness our proposed methods. The code at https://github.com/UCSD-AI4H/GraphSSL.