Enhancing Imbalanced Node Classification via Curriculum-Guided Feature Learning and Three-Stage Attention Network

TL;DR

This work tackles imbalanced node classification in graph neural networks by introducing CL3AN-GNN, a curriculum-guided feature learning framework that employs a three-stage attention mechanism (Engage-Enact-Embed). The model integrates modular, stage-dependent processing with adaptive curriculum loss to progressively handle easy and hard graph patterns, leveraging both structural and semantic neighbor information. Empirical results across eight Open Graph Benchmark datasets show consistent improvements in accuracy, F1-score, and AUC-ROC over strong baselines, along with analyses of gradient stability, attention correlations, and visualisations that support its effectiveness and interpretability. The approach advances practical imbalanced graph learning by delivering robust performance, faster convergence, and clear stage-wise explanations, with potential applicability to domains such as social, biological, and citation networks.

Abstract

Imbalanced node classification in graph neural networks (GNNs) happens when some labels are much more common than others, which causes the model to learn unfairly and perform badly on the less common classes. To solve this problem, we propose a Curriculum-Guided Feature Learning and Three-Stage Attention Network (CL3AN-GNN), a learning network that uses a three-step attention system (Engage, Enact, Embed) similar to how humans learn. The model begins by engaging with structurally simpler features, defined as (1) local neighbourhood patterns (1-hop), (2) low-degree node attributes, and (3) class-separable node pairs identified via initial graph convolutional networks and graph attention networks (GCN and GAT) embeddings. This foundation enables stable early learning despite label skew. The Enact stage then addresses complicated aspects: (1) connections that require multiple steps, (2) edges that connect different types of nodes, and (3) nodes at the edges of minority classes by using adjustable attention weights. Finally, Embed consolidates these features via iterative message passing and curriculum-aligned loss weighting. We evaluate CL3AN-GNN on eight Open Graph Benchmark datasets spanning social, biological, and citation networks. Experiments show consistent improvements across all datasets in accuracy, F1-score, and AUC over recent state-of-the-art methods. The model's step-by-step method works well with different types of graph datasets, showing quicker results than training everything at once, better performance on new, imbalanced graphs, and clear explanations of each step using gradient stability and attention correlation learning curves. This work provides both a theoretically grounded framework for curriculum learning in GNNs and practical evidence of its effectiveness against imbalances, validated through metrics, convergence speeds, and generalisation tests.

Figures (8)

• Figure 1: A typical progressive learning architecture with adaptive curriculum loss
• Figure 2: Overall proposed architecture of CL3AN-GNN model: The framework comprises four core modules: (A) The feature extraction module computes $\mathbf{h}_v^{(l+1)} = \sum_{u\in\mathcal{N}(v)}\text{Norm}(\mathbf{e}_{uv}^{(l+1)})$ to aggregate normalised neighbour features while $\mathbf{e}_{uv}^{(l+1)}$ capturing critical edge-node feature interactions. (B) The embedding layer processes edge features $\mathbf{e}_{uv}^{(l)}$ through multi-head attention to generate $\mathbf{z}_{vu}^{(l+1)} = \text{MultiHeadAttn}(W_e\mathbf{e}_{vu}, \mathbf{h}_v^{(l)}, \mathbf{h}_u^{(l)})$. (C) The curriculum learning module implements three-stage attention processing: the Engage stage computes initial attention via $W^{(l)}h_{uv}^{(l)}$, the Enact stage refines node and edge attention through $\alpha_{vu}^{(l+1)}$ and $\beta_{vu}^{(l+1)}$ respectively, and the Embed stage produces final attention scores $\mathbf{z}_{vu}^{(L)}$. (D) The classifier module generates predictions using these curriculum-refined features.
• Figure 3: Model evaluation on six datasets: Cora ($\lambda_1$), Citeseer ($\lambda_2$), PubMed ($\lambda_3$), Amazon Computers ($\beta_1$), Coauthor CS ($\beta_2$), and Amazon Photo ($\beta_3)$ based on AUC-ROC plots of True and False positive rates
• Figure 4: Model performance with AUC-ROC values for different oversampling scales on (a) Cora, (b) Citeseer, and (c) Chameleon datasets.
• Figure 5: Model attention stability correlations on OGBN-Arxiv dataset