Why Sacrifice Majority Nodes?: Improving Imbalanced Node Classification via Class-Balanced Graph Generation

ICLR 2026 Conference Submission15604 Authors

Keywords: imbalanced learning, imbalanced node classification

TL;DR: We propose a graph-generation approach for imbalanced node classification that overcomes a key limitation of existing methods.

Abstract: Class imbalance is prevalent in real-world data, often leading to a deterioration in a classifier’s generalization performance, especially on minority classes. Since graph-structured data is no exception, many efforts have been made to tackle imbalanced node classification by focusing on minority classes, leading to improved overall performance in imbalanced node classification. However, we find that these methods boost minority recall at the expense of degrading majority recall, a trade-off that has been overlooked. To address this issue, we propose Class Balancing Graph Generation (CBGG), a novel framework that prevents imbalanced node classifiers from sacrificing prediction power on majority classes. CBGG trains classifiers on high-quality synthetic graphs with class-balanced nodes, thereby tightening their generalization bounds across all classes. Extensive experimental results demonstrate that CBGG not only overcomes the majority-sacrifice pitfall of prior work but also significantly outperforms state-of-the-art imbalanced node classification methods across seven benchmark datasets.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 15604