Kernel-based learning in the absence of counterexamples: one-class classification

Abstract

This thesis mainly investigates the kernel learning-based approach for outlier (nov elty or anomaly or negative) detection using one-class classification (OCC). OCC is a non-traditional way of classification where the model is built using samples from only one class, and samples from this class belong to normal (positive or target) class. The one-class classifier classifies any unknown class other than the normal class as an outlier class. All proposed one-class classifiers in this thesis are developed based on the boundary and reconstruction frameworks. In recent years, kernel ridge regression (KRR) (or least squares support vector machine with zero bias or kernel extreme learning machine) based one-class classifiers have received quite an attention by researchers. Researcher developed a KRR-based one-class classifier for boundary framework. We have developed it for the recon struction framework. For further performance improvement, we have combined the concept of both the frameworks in a single multi-layer architecture. This architecture is formed by sequential stacking of various KRR-based Auto-Encoders, followed by a KRR-based one-class classifier. The stacked architecture provides a better repre sentation of the data using representation learning, which helps in obtaining better classification compared to single hidden layer-based architecture. Further, this multi layer architecture is extended to use structural information between samples using a Graph-Embedding approach. The structural information is generated by different types of Laplacian graphs and embedded into the existing multi-layer architecture. Later, we have explored multiple kernel learning (MKL) for one-class classification, which captures different notions of the data using different types of kernels.