Research map / Graph learning / GNN architectures

Cluster-guided Contrastive Class-imbalanced Graph Classification

AAAI 2025 · Graph learning › GNN architectures · approach: Graph structure learning · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes C$^3$GNN, which clusters majority-class graphs into size-balanced subclasses, applies Mixup within subclasses, and trains with hierarchical supervised contrastive losses to tackle class-imbalanced graph classification.

Problem

Graph classification under class-imbalanced training distributions, where GNNs become biased toward majority classes and existing vision class-imbalance methods overlook majority-class substructures and overfit minority classes.

Approach

Adaptively cluster each majority class into multiple subclasses sized comparably to the minority classes, enrich subclasses with Mixup interpolation, and apply intra-subclass and inter-subclass supervised contrastive losses to learn hierarchical graph representations.

Key results

• C$^3$GNN achieves 43.60% on Synthie (IF=15) vs. 42.35% for the second-best RAHNet.
• C$^3$GNN reaches 77.38% on MNIST (IF=50), outperforming RAHNet at 75.12%.
• Ablation shows removing hierarchical subclass contrastive learning causes the largest drop (e.g., MNIST 77.38 to 74.23).

Measured effects reported in the paper

Surprising findings

• Removing subclass mixup interpolation on Synthie slightly improves over the complete model variant in some ablation columns is not reported; instead, the w/o SMI variant on Synthie (43.08) is below the complete model (43.60).

Limitations

• Cluster size hyperparameter delta must be tuned per dataset
• Evaluated only on graph benchmarks with synthetic imbalance via Zipf-style splits

Experimental setup

Datasets: Synthie ENZYMES MNIST Letter-high Letter-low COIL-DEL

Benchmarks: Synthie ENZYMES MNIST Letter-high Letter-low COIL-DEL

Metrics: top-1 accuracy

Baselines: GraphSAGE Up-sampling CB loss LACE loss GraphCL SupCon Augmentation G$^2$GNN RAHNet

Builds on: supervised contrastive learning Mixup GraphCL G$^2$GNN

Closest papers by meaning

• ImGCL: Revisiting Graph Contrastive Learning on Imbalanced Node Classification (AAAI 2023 · similarity 0.92)
• Class-Imbalanced Graph Learning without Class Rebalancing (ICML 2024 · similarity 0.91)
• IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning (ICLR 2025 · similarity 0.88)
• PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering (ICML 2025 · similarity 0.87)
• Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
• ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.87)