Self-Optimizing Radial Basis Function Support Vector Classifier (SO-RBFSVC)

Abstract

Support vector classifiers (SVCs) typically use radial basis function (RBF) kernels to map data into higher dimensional spaces that may improve the linear separation of otherwise nonseparable classes. We present a novel self-optimizing radial basis function support vector classifier (SO-RBFSVC) that integrates response surface methodology (RSM), two-dimensional cubic spline interpolation, and bootstrapped Latin partitions (BLPs) for automated hyperparameter tuning. The SO-RBFSVC simultaneously optimizes the RBF kernel width (σ) and cost parameter (C) using an interpolated response surface obtained from generalized prediction accuracies. The SO-RBFSVC was compared to other self-optimizing classifiers (super SVC [sSVC] and super partial least squares discriminant analysis [sPLS-DA]). Four datasets were evaluated: (i) hemp and marijuana discrimination using proton nuclear magnetic resonance spectra, (ii) barley growth location using near-infrared spectra, (iii) glass-type identification based on elemental composition, and (iv) wine cultivar classification from physicochemical properties. External validation results showed that SO-RBFSVC performed comparably to the other models, achieving error rates of 0.4 ± 0.5% for hemp/marijuana, 7 ± 1% for glass, and 6 ± 1% for wine, while outperforming the linear models with 10 ± 1% error for the barley NIR data. For the first time, generalized sensitivity analysis (GSA) was applied to quantify model linearity. GSA revealed high nonlinearity in the barley dataset, justifying a nonlinear model. The SO-RBFSVC provides robust, automated classifier tuning for low- and high-dimensional datasets, offering ease of use.