Kernels for One-Class Nearest Neighbour Classification and Comparison of Chemical Spectral Data

Abstract

The One-class Classification (OCC) problem is different from the conventional binary / multi-class classification problem in the sense that in OCC, the examples in the negative / outlier class are either not present, very few in number, or not statistically representative of the negative concept. Researchers have addressed the task of OCC by using different methodologies in a variety of application domains. This thesis formulates a taxonomy with three main categories based on the way OCC is envisaged, implemented and applied by various researchers in different application domains. Based on the proposed taxonomy, we present a comprehensive research survey of the current state-of-the-art OCC algorithms, their importance, applications and limitations. The thesis explores the application domain of Raman spectroscopy and studies several similarity metrics to compare chemical spectra. We review some standard, non-standard and spectroscopy-specific spectral similarity measures. We also suggest a modified Euclidean metric to aid in effective spectral library search. These spectral similarity methods are then used to build the kernels for developing one-class nearest neighbour classifiers. Our results suggest that these new similarity measures indeed lead to better precision and recall rates of target spectra in comparison to studied standard methods. The thesis proposes the use of kernels as distance metric to formulate a one-class nearest neighbour approach for the identification of a chemical target substance in mixtures. The specific application considered is to detect the presence of chlorinated solvents in mixtures, although the approach is equally applicable for any form of spectral analysis. We use several kernels including polynomial (degree 1 and 2), radial basis function and spectral data specific kernels. Our results show that the radial basis function kernel consistently outperforms other kernels in one-class nearest neighbour setting. But the polynomial and spectral kernels perform no better than the linear kernel (which is directly equivalent to the standard Euclidean metric).