Nia-GNNs: neighbor-imbalanced aware graph neural networks for imbalanced node classification

It has been proven that Graph Neural Networks focus more on the majority class instances and ignore minority class instances when the class distribution is imbalanced. To address the class imbalance problems on graphs, most of the existing approaches rely on the availability of minority nodes in the training set, which may be scarce in extremely imbalanced situations and lead to overfitting. To tackle this issue, this paper proposes a novel oversampling-based Neighbor imbalanced-aware Graph Neural Networks, abbreviated as Nia-GNNs. Specifically, we propose a novel interpolation method that selects interpolated minority nodes from the entire dataset according to their predicted labels and similarity. Meanwhile, a class-wise interpolation ratio is applied to prevent the generation of out-of-domain nodes. Additionally, the generated minority nodes are inserted into the neighbor of minority nodes according to their neighbor distribution to balance the graph both neighborly and globally. Numerous experiments on different imbalanced datasets demonstrate the superiority of our method in classifying imbalanced nodes.