Development of Robust and Scalable Hyperbox based Machine Learning Algorithms
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Together with the rapid development of digital information and the increase in amount of data, machine learning (ML) algorithms have been developed and evolved constantly to discover new information and knowledge from different data sources. The use of hyperbox fuzzy sets as fundamental representational and building blocks in learning algorithms forms an important branch of ML. Hyperbox-based algorithms have a huge potential for high scalability and incremental adaptation to applications working in the dynamically changing environments. Additionally, learning algorithms based on hyperbox representations can form interpretable models, which are highly desirable for areas with the requirement of safety and trust. This study aims to develop and expand robust, scalable, and transparent learning algorithms for hyperbox-based classification models with a specific focus on a general fuzzy min-max neural network (GFMMNN).
First of all, a comprehensive survey on hyperbox-based machine learning models together with empirical assessments of the GFMMNN on pattern classification problems were conducted. Next, a new online learning algorithm was proposed for the GFMMNN and improved the robustness of the whole family of GFMMNN learning algorithms to work effectively with mixed-attribute data by introducing a new learning mechanism for categorical features. In terms of scalability, the main steps of the learning algorithms were reformulated so they can be effectively executed on graphics processing units using matrix operations, simultaneously proposing mathematical lemmas to reduce the redundancies of hyperbox candidates in the learning process. This thesis also proposed a novel method to enhance the transparency of classifiers while maintaining a good classification performance by using hierarchical granular representations from hyperbox fuzzy sets. The last contribution was a simple but powerful ensemble model built from many individual hyperbox-based classifiers trained on random subsets of both sample and feature spaces. Extensive empirical analyses indicated that the proposed solutions are highly competitive with other evaluated learning algorithms.
Please use this identifier to cite or link to this item: