Induction of Accurate and Interpretable Fuzzy Rules

Authors Organisations
Type

Student thesis: Doctoral ThesisDoctor of Philosophy

Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date2017
Links
Show download statistics
View graph of relations

Abstract

Knowledge discovery from data with fuzzy modelling is currently an active research area in the field of computational intelligence. Fuzzy modelling describes systems by establishing relationships between input and output variables with fuzzy logic and fuzzy set theory. One of the main advantages of fuzzy modelling lies in the interpretability, such that they can formulate the knowledge with linguistic fuzzy rules to gain insights into behaviours of a complex system. However, the interpretability is not automatically given due to only using fuzzy rules. Unlike accuracy that can be used to objectively assess performance of the underlying system, interpretability is a subjective property that may be affected by a range of practical issues, especially regarding the representation of the underlying concepts and domain knowledge. Despite of no commonly accepted mechanism to adjudge interpretability, the incorporation
of domain expertise encoded as predefined fuzzy sets is desirable to effectively
interpret a fuzzy model. This facilitates enhanced transparency in both learning the models and the inferences performed with the learned models.

In light of this, the thesis is focused on the automatic generation of accurate and
interpretable fuzzy models expressed as classification rules, where the use of fixed and predefined quantity spaces is a must for semantic interpretability. In this thesis, several approaches are presented with generated fuzzy rules being interpretable, and achieving competitive performance in comparison to state-of-the-art methods. These include: 1) the approach for the acquisition of fuzzy rules with quantifiers following class-dependent simultaneous rule learning strategy; 2) the approach for the acquisition of weighted fuzzy rules where heuristically generated fuzzy rules are initialised, followed by the global search of optimal rule weights; and 3) the approach that works by utilising existing crisp rules generated by a certain crisp rule-based learning classifier, and then performs rule mapping, followed by global genetic rule and condition selection. Furthermore, to enhance the capability of a fuzzy classifier, the thesis also develops a classifier ensemble approach based on the measure of nearest-neighbour-based reliability. Apart from benchmark data sets that have been utilised for systematic experimental verification, the proposed techniques
are applied to a real-world problem of academic journal ranking, demonstrating the efficacy of the present research.