A novel twin parametric-margin support vector machine with capped asymmetric elastic net loss.

Abstract

Support vector machine (SVM) is a widely used classifier, including hinge-SVM and twin parametric-margin support vector machine (TPWSVM). TPWSVM constructs two nonparallel hyperplanes by solving two smaller quadratic programming problems, demonstrating efficiency on large-scale datasets. However, conventional TPWSVM relies on hinge loss, leading to sensitivity to noise and instability under resampling. To overcome these drawbacks, we propose a novel capped asymmetric elastic net twin parametric-margin support vector machine (CaEN-TPMSVM), integrating capped asymmetric elastic net loss within the TPWSVM framework. Our method generalizes TPWSVM, improves noise robustness, and achieves a fourfold acceleration in training speed relative to standard SVM. Theoretical analysis demonstrates its convergence and stability properties. Empirical studies on synthetic and ten UCI datasets confirm its superior classification accuracy and computational efficiency.