NodeImport Improves Imbalanced Node Classification on Graphs

Node classification on graphs often suffers from class imbalance, where a few majority classes dominate training, leading to biased model performance. Traditional Graph Neural Networks (GNNs) typically struggle in these scenarios, overfitting to common classes and underrepresenting minority ones. Existing solutions, which either prioritize nodes by class size or synthesize new nodes, often fall short in effectively resolving this issue. This paper introduces NodeImport, a novel approach to tackle class-imbalanced node classification. The core idea is to utilize a balanced meta-set to measure node importance. A training node is deemed significant if its inclusion improves model performance in an unbiased setting. This method allows for fine-grained, dynamic selection of important nodes throughout the training process, counteracting the imbalance. NodeImport provides a theoretically derived formula for direct node importance assessment, reducing computational overhead and offering a clear selection threshold. The framework filters valuable labeled, unlabeled, and synthetic nodes, ensuring compatibility with various node generation methods by separating generation from filtering. It also includes a strategy for constructing a high-quality meta-set that accurately represents the overall feature distribution. Evaluated against popular GNNs and multiple datasets, NodeImport consistently outperforms existing baselines, demonstrating its flexibility and effectiveness in mitigating class imbalance.