Relabeling Noisy Labels: A Twin SVM Approach

In practical applications, supervised learning algorithms, including support vector machine (SVM), heavily rely on precise labeling to train predictive models. Nonetheless, real-world datasets often comprise mislabeled samples, which can have considerable influence on the performance of these algorithms. On the other hand, SVM suffers from computational costs when facing large-scale datasets. Twin support vector machine (TWSVM) tackles this issue and finds two nonparallel hyperplanes by solving two smaller models compared to SVM such that each one is closer to one of the two classes and is at least a unit distance far away from the samples of the other class. In this paper, to address label noise in datasets, we propose a TWSVM-based mixed-integer programming model for relabeling instances directly, while inheriting the advantages of TWSVM. Each model decides whether the samples of one class should be considered among instances that are as close as possible to its corresponding hyperplane. Therefore, each model exhibits the ability to recognize instances bearing close resemblance to one class while their assigned labels belong to the other one, prompting their reclassification. Conversely, instances demonstrating lower similarities to the other class retain their original labels. To show the efficiency of proposed models experiments are conducted on 12 UCI datasets.