Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.602707
Title: A study on imbalanced data classification problems
Author: Gao, Ming
Awarding Body: University of Reading
Current Institution: University of Reading
Date of Award: 2013
Availability of Full Text:
Access through EThOS:
Abstract:
This thesis focuses on the study of machine learning and pattern recognition algorithms for imbalanced data problems. The imbalanced problems are important as they are prevalent in life threatening/safety critical applications. They are known to be problematic to standard machine learning algorithms due to the imbalanced distribution between positive and negative classes. My original contribution to knowledge in this field is fourfold. A powerful and efficient algorithm for solving two-class imbalanced problems is proposed. The proposed method combines the synthetic minority over-sampling technique and the radial basis function classifier optimised by particle swarm optimization to enhance the classifier's performance for imbalanced learning. An over-sampling technique for imbalanced problems, probability density function estimation based over-sampling, is proposed. In contrast to existing over-sampling techniques that lack sufficient theoretical insights and justifications, the synthetic data samples are generated from the estimated probability density function from the positive data via the Parzen-window. A unified neurofuzzy modelling scheme is proposed. A novel initial rule construction method on the subspaces of the input features is formed. The supervised subspace orthogonal least square learning for model construction is applied. A logistic regression model is formed to present the classifiers output. Based on the formation of the unified neurofuzzy model, a new class of neurofuzzy construction algorithms is proposed with the aim of maximizing generalization capability specifically for imbalanced data classification based on leave-one-out cross-validation.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.602707  DOI: Not available
Share: