One-class classification algorithm

Abstract

One class classification (OCC) is a special case of multi-class classification where training data are exclusively derived from a single positive class. While conventional multi-class classification tasks as-sume availability of training data for all expected classes during prediction, OCC deals with scenarios where data from new or unforeseen classes emerge during testing. This project addresses the need for comprehensive comparisons of OCC algorithms, crucial for informed algorithm selection and advance-ment of anomaly detection methodologies. We introduce a selection of statistical and deep learning OCC methods and conduct a detailed analysis of their performance using image datasets. Specifically, we evaluate methods such as one class support vector machine (OCSVM), support vector data de-scriptor (SVDD), deep support vector data descriptor (DSVDD) and the holistic approach (HRN). Our analysis provides valuable insights into the strengths and limitations of OCC algorithms, facilitating their practical application in scenarios where obtaining labelled anomaly data is challenging. Through rigorous experimentation and comparison, we contribute to enriching understanding and guiding the selection of suitable OCC methodologies for diverse real-world applications.