Deep least squares one-class classification

Abstract

Listen

Translate article

One-class classification refers to a distinctive type of classification problem in which data originating from a single class, referred to as the target class, are exclusively available for training a specialized classifier known as the one-class classifier. This type of classification is encountered in various real-world applications when the characterization of outliers is challenging. Three primary methods—density estimation, boundary methods, and reconstruction methods—are used to solve one-class classification problems. This research is focused on boundary methods, particularly the popular support vector methods. Recently, the integration of deep neural networks with support vector methods has shown promise in one-class classification. This article proposes a new deep learning–based method called deep least squares support vector data description (DLS-SVDD), which combines the deep network’s feature extraction capability with the one-class objective to detect anomalies in complex data sets. The superior efficiency and performance of the proposed method over some one-class classifiers are illustrated through simulations and real-world data sets.