Resampling approach for one-Class classification
Hae-Hwan Lee, Seunghwan Park, Jongho Im
Abstract
The performance of a classification model depends significantly on the degree to which the support of each data class overlaps. Successfully distinguishing between classes is difficult if the support is similar. In the one-class classification (OCC) problem, wherein the data comprise only a single class, the classifier performance is significantly degraded if the population support of each class is similar. In this study, we propose a resampling algorithm that enhances classifier performance by utilizing the macro information that is most easily obtainable in these two problem situations. The algorithm aims to improve classifier performance by reprocessing the given data into data with mitigated class imbalance through raking and sampling techniques. This performance improvement is demonstrated by comparing representative classifiers used in the existing OCC problem with traditional binary classifier models, which are unavailable on a single-class dataset.