[论文解读] GraphSB: Boosting Imbalanced Node Classification on Graphs through Structural Balance

Zhixiao Wang, Chaofan Zhu+4 more|arXiv (Cornell University)|Jan 27, 2026

Advanced Graph Neural Networks被引用 0

一句话总结

tldr: GraphSB 引入结构平衡来解决图中固有的结构性不平衡，然后使用结构增强增强和扩散来提升基于 GNN 的节点分类中少数类的表现。

ABSTRACT

Imbalanced node classification is a critical challenge in graph learning, where most existing methods typically utilize Graph Neural Networks (GNNs) to learn node representations. These methods can be broadly categorized into the data-level and the algorithm-level. The former aims to synthesize minority-class nodes to mitigate quantity imbalance, while the latter tries to optimize the learning process to highlight minority classes. However, neither of them addresses the inherently imbalanced graph structure, which is a fundamental factor that incurs majority-class dominance and minority-class assimilation in GNNs. Our theoretical analysis further supports this critical insight. Therefore, we propose GraphSB (Graph Structural Balance), a novel framework that incorporates Structural Balance as a key strategy to address the underlying imbalanced graph structure before node synthesis. Structural Balance performs a two-stage structure optimization: Structure Enhancement that mines hard samples near decision boundaries through dual-view analysis and enhances connectivity for minority classes through adaptive augmentation, and Relation Diffusion that propagates the enhanced minority context while simultaneously capturing higher-order structural dependencies. Thus, GraphSB balances structural distribution before node synthesis, enabling more effective learning in GNNs. Extensive experiments demonstrate that GraphSB significantly outperforms the state-of-the-art methods. More importantly, the proposed Structural Balance can be seamlessly integrated into state-of-the-art methods as a simple plug-and-play module, increasing their accuracy by an average of 4.57%.

研究动机与目标

• Identify how intrinsic graph structure contributes to minority-class degradation in GNN-based node classification.
• Propose Structural Balance to mitigate structural imbalance via Structure Enhancement and Relation Diffusion.
• Show that Structural Balance improves minority-class learning and can be integrated into existing methods as a plug-and-play module.

提出的方法

• Theoretical analysis showing majority-class dominance and minority-class assimilation due to imbalanced graph structure.
• Two-stage Structural Balance: (1) Structure Enhancement to mine hard samples and strengthen minority connectivity; (2) Relation Diffusion to propagate enhanced minority context and model higher-order structure.
• Hard sample mining via dual-view analysis (feature view vs neighbor view) to detect minority-conflicting samples.
• Adaptive augmentation by adding edges from minority samples to minority anchors with a similarity constraint to preserve homophily.
• Sparse iterative diffusion on augmented graph with stochastic structure perturbation to capture higher-order dependencies with linear complexity.”
• Integration with a GNN classifier that combines original and synthesized information for training with a joint node+edge loss

实验结果

研究问题

• RQ1 How does imbalanced graph structure contribute to minority-class degradation in GNNs?
• RQ2 Can a structural intervention, before node synthesis, improve minority-class representation and classification accuracy?
• RQ3 Is Structural Balance compatible as a plug-and-play module with existing GNN-based imbalance methods?

主要发现

• GraphSB consistently outperforms state-of-the-art baselines across eight datasets in accuracy and Macro-F1.
• Structural Balance improves learning by addressing underlying structural imbalance, not just data-level or algorithm-level bias.
• SB can be integrated into other methods to improve accuracy by an average of 4.57%.
• Ablation shows Structure Enhancement and Relation Diffusion both contribute to performance, with SE particularly beneficial on severely imbalanced graphs.
• On large-scale graphs, SB yields substantial absolute gains for classical baselines, while converging performance among many methods when SB is applied.
• t-SNE visualizations indicate GraphSB produces clearer, well-separated class clusters compared to baselines.