CL3AN-GNN: curriculum-driven for incremental feature disentanglement with multi-phase attention for class-imbalanced node classification

Abstract

Imbalanced node classification in graph neural networks (GNNs) presents a significant challenge, where skewed class distributions lead to biased models that underperform on minority classes. To address this, we propose CL3AN-GNN , a curriculum guided three stage attention network that progressively disentangles features from class agnostic to class specific representations through structured learning phases: Engage , Enact , and Embed . The model begins by Engaging with structurally simpler nodes, characterised by low structural marginality and high feature purity, using uniform attention to build robust foundational representations. In the Enact stage, attention is relaxed to dynamically reweight neighbourhoods and enforce orthogonality between structural and semantic features, initiating feature disentanglement. Finally, the Embed stage specialises attention toward minority class nodes and hard examples, refining discriminative features with adaptive loss weighting. We evaluate CL3AN-GNN on eight imbalanced graph benchmarks spanning social, biological, and citation networks. Experiments demonstrate consistent improvements in accuracy, macro F1, and area under the curve over state-of-the-art methods, with particular gains on minority classes. The curriculum framework yields faster convergence, better generalisation to unseen imbalanced graphs, and improved interpretability via gradient stability analysis and attention correlation maps. This work provides a theoretically grounded and empirically validated approach to imbalanced node classification, advancing curriculum learning in graph representation learning.