Power-GNN: a graph over-sampling method to mitigate power-law distribution in graph neural networks

Since the advent of Graph Neural Networks (GNNs), they have been widely applied in the analysis and processing of graph data, especially demonstrating outstanding performance in semi-supervised node classification tasks. However, the class distribution in real-world graph data often exhibits a long-tail, imbalanced distribution, posing significant challenges to the classification performance of GNNs. Graph over-sampling methods address this by synthesizing new nodes for minority classes and creating corresponding edges, thus aiming to balance class representation and enhance model accuracy. Nonetheless, the degree distribution of nodes in reality also follows a power-law distribution, leading to synthesized nodes becoming low-degree tail nodes under existing edge construction strategies. This restricts their ability to acquire sufficient aggregation information, thereby degrading their representation quality and impacting classification outcomes. To address these challenges, this paper introduces Power-GNN, a novel graph data over-sampling framework tailored to tackle the dual challenges of imbalanced class distribution and the power-law distribution of node degrees. Power-GNN innovatively utilizes the power-law distribution of node degrees in a reverse manner. It strategically adds edges with high similarity to nodes with fewer connections, thereby amplifying the aggregation capability of synthesized nodes and boosting overall model performance. Through evaluations on multiple public benchmark datasets, Power-GNN has demonstrated superior performance over existing baselines across three common GNN architectures.