An Adaptive Robust Support Vector Machine With Sequential Minimal Optimization for Imbalanced Data Classification

Class imbalanced classification presents a considerable difficulty in machine learning, as conventional algorithms typically exhibit bias towards the majority class, compromising minority class recognition. Although a Robust Support Vector Machine (SVM) mitigates this issue through a fixed margin adjustment factor (μ), its static nature limits adaptability across varying imbalance levels. This study proposes an Adaptive Robust SVM (ARSVM) framework that introduces a dynamically computed adjustment factor (μ*) derived directly from the class imbalance ratio. The model was optimized via an enhanced Sequential Minimal Optimization (SMO) algorithm, integrating μ* into the Karush-Kuhn-Tucker (KKT) conditions and error computation. Extensive simulations across multiple imbalance ratios (0.007-0.5), sample sizes (1000-5000), and kernel functions (Radial Basis Function (RBF), Polynomial, Sigmoid) demonstrated that ARSVM consistently achieves superior balanced accuracy and AUC-ROC, particularly when paired with the RBF kernel in moderately imbalanced to balanced scenarios. In extreme imbalance settings, ARSVM with Polynomial kernel retains minority class detection capability when standard kernels fail. The results confirm that ARSVM effectively mitigates majority-class bias and enhances classification robustness by establishing a data-informed kernel selection framework for imbalanced learning. These findings offer both methodological advancement and practical guidance for real-world applications that require reliable performance under class distribution skew.