Doctorate thesis defense of Soumaya Nheri

Doctorate thesis defense on December 13^th 2023 at 13H00 AM ,in Amphitheater Ibn Khaldoun, SUP'COM 2.

Entitled :A Novel Subclass One-Class Classification for Pattern Recognition

Presented by :Soumaya Nheri

Committee

Abstract

In real-world scenarios such as nuclear power plants or rare medical diseases, gathering sufficient data on faulty or abnormal behavior is often challenging, while information from normal operations is more readily available during the training process. This imbalance necessitates the use of One-Class Classification (OCC) techniques to detect and identify abnormal instances. Various methods and concrete models have been proposed and developed to address OCC problems. These approaches aim to effectively distinguish between normal and abnormal behavior, enabling accurate detection and identification of faults or anomalies in real-world systems.

However, state-of-the-arts models do not consider low variance direction in order to improve classification accuracy. Thus, Covariance-guided One-Class Support Vector Machine (COSVM) enhances one-class classifiers by projecting data into low variance direction, while classifying data, thereby resulting in better classification performance. Nevertheless, COSVM struggles with high spread data or data with high dispersion.

This research introduces two novel models, namely, Dispersion COSVM (DCOSVM) and Scatter-Covariance OSVM (SC-OSVM) which address these limitations, in order to handle high spread data or data with high dispersion, in a proper manner. More precisely, the DCOSVM identifies data subclasses in the Kernel space, and projects them into the low variance directions, while performing classification. The SC-OSVM incorporates subclass information within the COSVM objective function, thereby exploiting subclass information in the kernel space. This allows a joint minimization of dispersion within and between subclasses, while classifying data. Both models enhance one-class classifiers and provide classification performance overhead.

Validation of the proposed models in handwritten digit recognition and human face detection has shown clearly their superiority and efficiency, in terms of classification performance.

Keywords

subclass learning, one-class classification, low variance direction, handwritten digit recognition, human face detection.