RoBell-RVFL Boosts Robustness in Imbalanced Data Learning

In real-world datasets, the prevalence of majority classes often biases randomized neural networks, causing them to overlook critical minority samples. While methods like SMOTE and class-weighted loss functions attempt to balance class proportions, they frequently neglect intra-class distribution, leaving them vulnerable to label noise and outliers. This paper introduces RoBell-RVFL, a robust and lightweight generalized bell random vector functional link network designed to tackle these challenges. RoBell-RVFL employs a dual-strategy, sample-level weighting mechanism. It strictly preserves information from minority classes by assigning unit weights. Simultaneously, it adaptively regulates the influence of majority class samples using a probability-weighted generalized bell (gbell) membership function within a kernel-induced feature space. This innovative design effectively suppresses noisy, boundary, and outlier samples within the majority class, enabling the network to learn from truly informative data points rather than merely abundant ones. By explicitly incorporating local class probability and class distribution information, RoBell-RVFL achieves adaptive control over sample contributions without sacrificing the computational efficiency inherent to RVFL networks. Extensive evaluations on UCI and KEEL benchmark datasets, including robustness tests with up to 40% label noise, demonstrate that RoBell-RVFL consistently and significantly outperforms recent state-of-the-art RVFL variants, proving the importance of quality-aware sample weighting in noisy and imbalanced environments.