GraphGS: Mitigating Pseudo-Label Shift in Imbalanced Node Classification via Balanced Feature Propagation and Minority Node Selection

Abstract

Graph-structured data in real-world applications often grapple with class imbalance, where underrepresented minority nodes result in biased predictions and diminished learning performance. Although pseudo-labeling methods offer a promising solution for addressing class imbalance, they remain susceptible to pseudo-label shift. This phenomenon occurs when class biases in pseudo-labels propagate during data augmentation, further compromising model effectiveness. To overcome these challenges, we present GraphGS, an innovative framework designed to alleviate pseudo-label shift in class-imbalanced graph datasets. GraphGS integrates two pivotal components: a feature balancer that curtails excessive information transfer from majority to minority nodes, safeguarding the unique characteristics of minority classes, and a node selector that identifies high-confidence minority nodes through similarity metrics. Together, these components enhance minority class representations without introducing noise, enabling more equitable and reliable training of Graph Neural Networks (GNNs). Comprehensive experiments on diverse real-world datasets underscore the robustness and adaptability of GraphGS. Notably, on the CiteSeer dataset, GraphGS achieves an outstanding 2.60% improvement in accuracy over state-of-the-art methods, marking a significant advance in minority class classification.