GraphCS: Community-Enhanced Minority Hard Point Generation Method

In graph classification tasks, the issue of class imbalance may lead models to be biased toward dominant classes, thereby reducing the recognition ability for minority classes and negatively impacting overall performance. To address this issue, this study proposes a community-enhanced minority hard point generation method. The method identifies "hard point" nodes in minority classes based on community strength, expanding the decision boundary to tackle the problem of spatial compression. Furthermore, graph diffusion smoothing is utilized to convert unweighted graphs into weighted ones, preserving the neighborhood structure's integrity. A differential graph diffusion strategy is adopted to generate the diffusion matrix. Experimental results demonstrate that this method significantly enhances the classification accuracy of Graph Neural Networks (GNNs) for minority class nodes on datasets such as Cora and Amazon-Photo, showcasing its effectiveness in solving class imbalance problems.