• Combines cluster mining with neighborhood-aware sampling so minority classes receive informative gradients.
• Produces up to 7% macro-F1 improvements across Open Graph Benchmark datasets.