Critical Review for One-class Classification: recent advances and the reality behind them

This paper presents a critical review of one‐class classification (OCC). Old articles defined OCC in a vague way, which allowed OCC models to learn from multiple classes. This paper reconsiders the OCC definition, as training data includes solely one class, and samples belonging to other classes are not available. Moreover, the review introduces a new OCC taxonomy consisting of boundary, distance, probability, fake, and subtask‐based approaches. Additionally, the article reveals that many OCC algorithms have learned multiple classes. Common violations include accessing unlabeled datasets, importing other datasets, and hyperparameter tuning based on the testing results. In addition, this paper suggests two gray zones in OCC: creating fake datasets and fake OCC problems from scratch, and decomposing samples into smaller units for accessing multiple classes. These gray zones could contribute to future theory to learn from a single class. On the other hand, the application of OCC can use multiple classes; generally, multiple classes outperform a single class. However, the applications will no longer be OCC after learning multiple classes.