Boosting RVFL networks: An outlier-resistant learning framework

The presence of noise and outliers in real-world data often degrades the reliability of classification and regression models. Traditional approaches, including fuzzy concepts and weighting schemes, often rely on specific function choices, which may limit their effectiveness. To address these challenges, we propose an AdaBoost-based random vector functional link network with bounded loss (AD-RVFL), which combines the adaptive reweighting capability of AdaBoost with the computational efficiency of RVFL networks. Unlike the traditional exponential loss used in AdaBoost, which can amplify the influence of outliers, the proposed AD-RVFL uses a bounded exponential loss that reduces the impact of extreme samples. For regression tasks, we further develop AD-RVFL-R, which reformulates the boosting strategy by introducing a relative-error-based thresholding mechanism while retaining analytically trained RVFL base learners. In addition, to analyze the impact of exponential loss in AdaBoost-based RVFL models, we present a classical AdaBoost RVFL model (CAD-RVFL), which employs the standard exponential loss. Experimental results and statistical analysis on several real-world datasets demonstrate that the proposed AD-RVFL and AD-RVFL-R achieve improved performance compared to baseline models. Furthermore, experiments under varying levels of outlier contamination show that AD-RVFL maintains strong robustness to outliers, highlighting its effectiveness.