Efficient and Effective Augmentation Framework With Latent Mixup and Label-Guided Contrastive Learning for Graph Classification

Graph Neural Networks (GNNs) with data augmentation obtain promising results among existing solutions for graph classification. Mixup-based augmentation methods for graph classification have already achieved state-of-the-art performance. However, existing mixup-based augmentation methods either operate in the input space and thus face the challenge of balancing efficiency and accuracy, or directly conduct mixup in the latent space without similarity guarantee, thus leading to lacking semantic validity and limited performance. To address these limitations, this paper proposes $\mathcal {G}$ -MixCon, a novel framework leveraging the strengths of Mix up-based augmentation and supervised Con trastive learning (SCL). To the best of our knowledge, this is the first attempt to develop an SCL-based approach for learning graph representations. Specifically, the mixup-based strategy within the latent space named $GDA_{gl}$ and $GDA_{nl}$ are proposed, which efficiently conduct linear interpolation between views of the node or graph level. Furthermore, we design a dual-objective loss function named SupMixCon that can consider both the consistency among graphs and the distances between the original and augmented graph. SupMixCon can guide the training process for SCL in $\mathcal {G}$ -MixCon while achieving a similarity guarantee. Comprehensive experiments are conducted on various real-world datasets, the results show that $\mathcal {G}$ -MixCon demonstrably enhances performance, achieving an average accuracy increment of 6.24%, and significantly increases the robustness of GNNs against noisy labels.