p–LSGR: A lightweight post–hoc latent–space resampling framework for imbalanced and complex graph distributions

Graph neural networks (GNNs) are widely used across domains but remain sensitive to class imbalance, class overlap, and complex data distributions, limiting their reliability in real-world settings. Existing imbalance-mitigation strategies provide only partial robustness, are often computationally expensive, and post-hoc imbalance handling remains underexplored in graph learning. We propose Posthoc Latent Space Graph Resampling (p-LSGR), a framework that exploits learned graph representations through latent-space resampling to improve minority-class learning without graph augmentation or modification of the original graph structure. By integrating data-aware, difficulty-aware, and ensemble-based resampling strategies, p-LSGR leverages learned representations to improve minority-class classification performance. We evaluate p-LSGR across diverse graph datasets, comparing against pre-training and in-training imbalance-mitigation methods. On average, p-LSGR achieves a 42% improvement in minority-class F1 over vanilla GNNs and a 6% improvement over competing approaches while reducing processing time by up to 91%. Bayesian analysis further indicates that p-LSGR variants maintain strong minority-class performance while achieving balanced tradeoffs across Accuracy and AUPRC. These findings position p-LSGR as a practical and computationally efficient alternative to graph-level augmentation. We additionally release a corpus of constraint-based synthetic graph datasets derived from real electrical layouts to support research on class-imbalanced graph learning at https://data.mendeley.com/datasets/4vdpxdd7vp/1