Towards Equitable Node Classification: A Unified Graph Neural Network Framework
Chun Hing Cheng, Cuicui Luo · 2024
Node classification is a critical research area within the realm of graph neural networks. Many real-world node classification scenarios exhibit significant imbalances among nodes, causing graph neural networks to heavily favor instances from the majority class. While existing methods for addressing class imbalance in other domains can partially alleviate this issue, they often overlook the impact of graph structure. To tackle overfitting during the sampling process, this study introduces a unified graph neural network model called ESMOTE4Graph, tailored for fair oversampling. The model comprises four key modules: a feature extractor, synthetic node generator, edge generator, and a classifier with a class-balanced optimization objective. By introducing randomness during the oversampling phase, the model aims to generate diverse nodes. It also employs rigorous feature and neighbor selection mechanisms to ensure that the generated nodes retain the characteristics and neighborhood distribution of their respective classes. Utilizing a class-balancing loss function, the model adjusts the decision boundary of the graph neural network to achieve optimal performance. To evaluate the model's efficacy, we conducted experiments on multiple datasets using various baseline graph neural network models. Our experimental findings demonstrate that ESMOTE4Graph surpasses other models in the task of imbalanced node classification.