noisy label relabeling by nonparallel support vector machine

In machine learning, models are derived from labeled training data where labels signify classes and features define sample attributes. however, noise from data collection can impair the algorithm’s performance. blanco, japón, and puerto proposed mixed-integer programming (mip) models within support vector machines (svm) to handle label noise in training datasets. nonetheless, it is imperative to underscore that their models demonstrate an observable escalation in the number of variables as sample size increases. the nonparallel support vector machine (npsvm) is a binary classification method that merges the strengths of both svm and twin svm. it accomplishes this by determining two nonparallel hyperplanes by solving two optimization problems. each hyperplane is strategically positioned to be closer to one of the classes while maximizing its distance from the other class. in this paper, to take advantage of npsvm’s features, npsvm-based relabeling (renpsvm) mip models are developed to deal with the label noises in the dataset. the proposed model adjusts observation labels and seeks optimal solutions while minimizing computational costs by selectively focusing on class-relevant observations within an ϵ-intensive tube. instances exhibiting similarities to the other class are excluded from this ϵ-intensive tube. experiments on 10 uci datasets show that the proposed npsvm-based mip models outperform their counter parts in accuracy and learning time on the majority of datasets.