Use this URL to cite or link to this record in EThOS:
Title: Evolutionary induction of projection maps for feature extraction
Author: Rodriguez Martinez, Eduardo
ISNI:       0000 0004 2737 3755
Awarding Body: University of Liverpool
Current Institution: University of Liverpool
Date of Award: 2011
Availability of Full Text:
Access from EThOS:
Access from Institution:
This thesis proposes an evolutionary scheme for automatic design of feature extraction methods, tailored to a given classification problem. The main advantage of the proposed scheme is its capacity to formulate new models when the exiR.tmg'ones do not fit the .. ' problem at hand. The learning phase is expressed as a model selection problem, where the best performing model is selected among the genetic pool, assessed by an estimation of out-of-sample generalization error. Each individual in the genetic pool represents a potential model encoded into a hybrid genotype, specifically designed to hold a tree structure and an scalar array to represent both feature-extraction and classification stages. The role of the inducer is to automatically design a mapping function to be used as the core of the feature-extraction stage, as well as fine-tune the corresponding hyper-parameters for the feature-extraction/classification pair. Two paradigms are explored to express the feature-extraction stage, namely projection pursuit and spectral embedding methods. Both paradigms can express several feature extraction algorithms under a common template. In the case of projection pursuit, such template consist on the optimisation of a cost function, also known as projection index, that can be specifically designed to highlight certain properties of the extracted features. While for spectral embedding methods, a suitable set of similarity metrics is needed to construct a weight matrix, which encodes the links between any two samples on the vertices of a graph. The eigendecomposition of such weight matrix represents the solution to an optimisation problem looking for a low-dimensional space, retaining the characteristics described by the original distance metric. The proposed inducer evolves an optimal projection index or a desired distance metric for the corresponding feature-extraction paradigm. Addi- tionally, projection pursuit was extended to the nonlinear case by means of the kernel trick. The determination of a nonlinear residual subspace for sequential projection pursuit is reduced to the computation of an updated kernel matrix.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral