Less is More: Using Buffer Nodes to Reduce Excessive Majority Node Influence in Class Imbalance Graphs

Qian Wang, Zemin Liu, Zhen Zhang, Bingqiao Luo, Bingsheng He

Keywords: Class imbalance

Abstract: Graph Neural Networks (GNNs), despite success in node classification, struggle with class-imbalanced graphs, leading to minority node misclassification. Existing methods that synthesize minority nodes often overlook how majority nodes propagate misleading information through majority-minority edges; our analysis confirms this negative impact. To address this, we propose BufferGraph, a framework that inserts buffer nodes on such edges. These nodes act as controlled bottlenecks to reduce excessive majority node influence. And we theoretically demonstrate they reduce minority node feature distortion. Experiments on five real-world datasets show BufferGraph improves accuracy by up to 2\% over state-of-the-art methods, excelling in imbalanced settings and for minority classes with high heterophily. Code is available at \url{https://anonymous.4open.science/r/BufferGraph-C257}.

Submission Type: Full paper proceedings track submission (max 9 main pages).

Submission Number: 135