Inter-Class and Intra-Class Re-Weighting for Topological-Imbalance Graph Neural Networks

Gen Liu, Zhongying Zhao, Hui Zhou, Chao Li, Qingtian Zeng, Guanfeng Liu

Abstract: Graph Neural Networks (GNNs) are powerful tools for solving node classification on graphs. In the task of semi-supervised node classification, the existing studies often assign an equal number of labeled nodes for each class to maintain quantity balance and avoid the long-tailed issue. Once falling into such a situation, the existing methods typically adopt class re-balancing strategies such as re-sampling and re-weighting based on the number of observations for each class. However, the issue of topological imbalance remains persistent within the graphs. In this paper, we first analyze the problem of topological imbalance in terms of inter-class and intra-class perspectives. We then propose an inter-class and intra-class re-weighting method, named I2RW, to alleviate the influence of topological imbalance in semi-supervised node classification. Specifically, the inter-class re-weighting method is designed to recalibrate the importance of various classes based on the loss distribution. Furthermore, the intra-class re-weighting method is developed to fine-tune the influence of nodes within the same class based on their respective node degree information. Note that the proposed I2RW can be combined with existing GNN-based models in a plug-and-play way. Experimental results on five real-world datasets demonstrate that our proposed I2RW yields notable performance improvements when applied to various GNN-based models.

External IDs:dblp:journals/tetci/LiuZZLZL25