Geometric one-class classifiers using hyper-rectangles for knowledge extraction

Abstract

One-class classification (OCC) is to properly classify unknown data by developing a classifier, which can learn about a given training dataset with only one class. Its importance has been increasing in situations such that (i) there exists little or no information pertaining to other classes, or (ii) there are many unknown and heterogeneous classes. However, many OCC algorithms developed so far have some limitations as (i) they are black-box algorithms so that the user cannot clearly understand the internal classification mechanism of generated models, and (ii) there are no dominant criteria for node branching in decision-tree-based algorithms. Based on these motivations, in this work, we propose two efficient one-class classifiers using hyper-rectangles (h-rtgl) to describe a dataset with only one class: one-class hyper-rectangle descriptor by merging and one-class hyper-rectangle descriptor by partitioning, depending on the methods generating intervals and h-rtgls. The suggested classifiers can control the number and volume of the generated h-rtgls, which can affect the classification accuracy and deal with overfitting issue. We show the superiority of the proposed classifiers by a numerical experiment using UCI machine learning repository.