A kernel SVM algorithm to detect mislabeled microarrays in human cancer samples

DNA Microarrays have been successfully applied to the identification of different cancer types considering the gene expression profiles. However, previous studies have shown that labeling errors are not uncommon in microarray studies. In this case, the training set may contain mislabelled examples that may lead the classifier to poor performance. In this paper we propose a new filtering algorithm based on one-class SVM classification to detect mislabelled samples. To this aim, samples and labels are mapped together to feature space using the kernel of dissimilarities. Next, outliers are detected via one-class classification. Mislabeled samples and outliers in input space can be separated comparing the outliers obtained in input and feature spaces. The algorithm proposed has been tested using several complex cancer microarray datasets in which some samples are mislabelled according to the literature. The experimental results suggest that our algorithm is effective detecting labeling errors and compares favorably with a standard technique such as simple SVM.