Imbalanced Graph Classification based on Graph-of-Graph Neural Networks and Augmented Graph Generator
Minxi Rong, Qi Sun, Guo Xiao-li, Guipeng Zhang, Huijing Qi, Fuyang Hu · 2025
Graph classification has significant applications in various domains, including chemical molecular structure prediction, social network analysis, recommendation systems, and biological network analysis, and has received considerable attention in recent years. However, real-world graph datasets often exhibit class imbalance, where the disproportionate distribution of classes can cause models to favor the majority class, reducing prediction accuracy for minority classes. Notably, minority or anomalous classes often hold unique value in many practical scenarios. This paper proposes a framework for graph classification under category imbalance named Generator Imbalanced Graph Classification(GIGC). GIGC employs a set of learnable augmented graph generators orchestrated by an automated augmentation strategy to enhance minority-class graphs. This approach increases the diversity of minority-class samples, balances the dataset's class distribution, and improves the model's generalization capabilities. Simultaneously, k-reciprocal nearest neighbors are utilized to capture inter-graph relationships more effectively, leveraging the properties of symmetry and reciprocity. Subsequently, Graph-of-Graph Neural Networks are used to model inter-graph relationships, capturing global associations between graphs at a higher level. This enables the model to better capture and utilize inter-graph information. By leveraging the latent connections between minority-class graphs and other graphs, GIGC enriches the representation of minority-class graphs, mitigating the negative effects of class imbalance on classification performance. Extensive experiments on five real-world datasets demonstrate that GIGC outperforms baseline methods in graph classification. Specifically, F1-macro and F1-micro are improved by at least 1.12% and 2.62%, respectively. These results validate the effectiveness of GIGC in addressing class imbalance in graph classification tasks.