ABSTRACT One‐class classification presents a unique classification problem aimed at addressing the intricate challenges associated with detecting anomalies within datasets that do not exhibit known outliers. With a concentrated focus on modeling a single class of data, this approach naturally aligns with the core objectives inherent in anomaly detection. The task of discerning anomalies within datasets predominantly composed of normal instances holds critical importance, particularly in fields such as healthcare, finance, and manufacturing, where the characterization of outliers is notably challenging. This paper proposes Deep Least Squares One‐Class Support Vector Machine (DLS‐OCSVM), a one‐class classification method tailored specifically for the detection of rare and unusual events in complex data. DLS‐OCSVM strategically combines the strength of neural networks in learning rich feature representations with the well‐established One‐Class Support Vector Machine (OCSVM), using a least squares one‐class objective. The effectiveness of the proposed method is demonstrated through simulation studies and real‐world data. We compare the performance of DLS‐OCSVM with several established one‐class classification methods, including the kernel‐based OCSVM, Least Squares One‐Class Support Vector Machine (LS‐OCSVM), and Support Vector Data Description (SVDD), the deep learning methods One‐Class Neural Network (OC‐NN) and Deep Support Vector Data Description (DSVDD), and Isolation Forest (iForest). The proposed one‐class classifier demonstrates competitive performance comparable to existing deep learning methods, while outperforming the kernel‐based methods under the settings considered.
Hampton et al. (Wed,) studied this question.