Twin support vector machine is sensitive to noise. Existing weighted variants such as the fuzzy twin support vector machine characterize samples only along single or partial dimensions and provide limited ability to address feature redundancy.
This paper proposes a typicality measure and outcome-adaptive twin support vector machine (TATSVM), which integrates typicality-based sample weighting with an outcome-adaptive mechanism. At the sample level, a local density-margin contrast typicality is constructed to better distinguish informative and noisy samples. At the feature level, outcome-adaptive $L_2$ regularization for linear models and an outcome-adaptive distance metric learning-based Gaussian kernel for nonlinear models enable dynamic suppression of redundant features.
Experiments on 2 synthetic datasets and 10 UCI datasets demonstrate that TATSVM achieves higher accuracy, exhibits smaller degradation under noise and achieves a better trade-off among accuracy, robustness and computational efficiency.
By denoising at the sample level and optimizing at the feature level, the model can simultaneously address sample noise and high-dimensional redundancy to significantly enhance robustness.
