Linear and ellipsoidal pattern separation: theoretical aspects and experimental analysis

Andriy Kharechko

    Research output: Book/ReportCommissioned reportpeer-review

    Abstract

    This thesis deals with a pattern classification problem, which geometrically implies data separation in some Euclidean feature space. The task is to infer a classifier (a separating surface) from a set or sequence of observations. This classifier would later be used to discern observations of different types. In this work, the classification problem is viewed from the perspective of the optimization theory: we suggest an optimization problem for the learning model and adapt optimization algorithms for this problem to solve the learning problem. The aim of this research is twofold, so this thesis can be split into two self-contained parts because it deals with two different type of classifiers each in a different learning setting. The first part deals with linear classification in the online learning setting and includes analysis of existing polynomial-time algorithms: the ellipsoid algorithm and the perceptron rescaling algorithm. We establish that they are based on different types of the same space dilation technique, and derive the parametric version of the latter algorithm, which allows to improve its complexity bound and exploit some extra information about the problem. We also interpret some results from the information-based complexity theory to the optimization model to suggest tight lower bounds on the learning complexity of this family of problems. To conclude this study, we experimentally test both algorithms on the positive semidefinite constraint satisfaction problem. Numerical results confirm our conjectures on the behaviour of the algorithms when the dimension of the problem grows. In the second part, we shift our focus from linear to ellipsoidal classifiers, which form a subset of second-order decision surfaces, and tackle a pattern separation problem with two concentric ellipsoids where the inner encloses one class (which is normally our class of interest, if we have one) and the outer excludes inputs of the other class(es). The classification problem leads to semidefinite program, which allows us to harness the efficient interior-point algorithms for solving it. This part includes analysis of the maximal separation ratio algorithm
    Original languageUndefined
    Publication statusPublished (in print/issue) - 1 Jul 2009

    Cite this