Clustering-based proxy measure for optimizing one-class classifiers

Abstract

One-class classification is a type of unsupervised learning task wherein only information on the target class is available and that on other classes is not. As this corresponds to many real-world data scenarios, considerable research efforts have been focused on developing one-class classification algorithms to build well-performing one-class classifiers. The performance of one-class classifiers largely depends on their hyperparameters. However, optimizing hyperparameters is inherently difficult and challenging because conventional validation procedures are not directly applicable in the absence of non-target class information. The existing methods have focused on specific one-class classification algorithms, and therefore, could not be extended to other algorithms. In this study, we present a clustering-based proxy measure to evaluate the quality of different hyperparameter candidates for a one-class classification algorithm when only the target class information is given. The proposed proxy measure utilizes conventional measures of classification performance by introducing pseudo non-target classes that are derived based on clustering. The best hyperparameters for the algorithm can be estimated by alternatively optimizing this measure. It does not require incorporating any prior knowledge, and can be implemented with various types of one-class classification algorithms. We demonstrate the effectiveness through experiments comprising benchmark problems.