Label Contrastive Coding Based Graph Neural Network for Graph Classification

Graph classification is a critical research problem in many applications from different domains. In order to learn graph model, the most widely used supervision component an output layer together with loss (e.g., cross-entropy softmax or margin loss). fact, discriminative information among instances are more fine-grained, which can benefit tasks. this paper, we propose novel Label Contrastive Coding based Neural Network (LCGNN) utilize label effectively and comprehensively. LCGNN still uses ensure discriminability of classes. Meanwhile, leverages proposed Loss derived self-supervised learning encourage instance-level intra-class compactness inter-class separability. To power contrastive learning, introduces dynamic memory bank momentum updated encoder. Our extensive evaluations eight benchmark datasets demonstrate that outperform state-of-the-art models. Experimental results also verify achieve competitive performance less training data because exploits