Automated Multi-scale Contrastive Learning with Sample-Awareness for Graph Classification

Proper sample selection can better facilitate mutual information learning. Current sample selection methods suffer from fragile circularity, dependence on labeling information, and an imbalance in the volume of sample information. To address these challenges, we propose an automated multi-scale contrastive learning with sample-awareness for graph classification approach (SaMGCL), which realizes deep coupling of feature learning and sample selection through graph network science. More precisely, SaMGCL devises automated augmentors to fuse the complex features of nodes and edges, thereby enhancing the richness of sample information. Different levels of representations are optimized interactively through an end-to-end multi-scale contrastive learning framework. Especially, considering the actual graph structural noise and node heterogeneity, we extract high-quality samples oriented to the topology of the input graph and refine neighborhood information. Extensive experiments on eight benchmark datasets demonstrate that our proposed SaMGCL achieves superior graph classification performance compared to the current state-of-the-art approaches.