Evaluation of fuzzy membership functions for linguistic rule-based classifier focused on explainability, interpretability and reliability

Objective: Decision support systems focus on their interpretability with a different caution. The majority of approaches utilize reliable optimization techniques to achieve high recognition accuracy. However, accuracy improvement is done at the cost of the readability of the decision support system. This paper proposes a novel technique of linguistic rule extraction focused on: explainability, interpretability, and reliability, called EasIeR. Methods: Proposed solution performs a direct analysis of the learning set without omitting real data imperfections. The proposed algorithm processes the imprecision of data with linguistic values described by fuzzy membership functions and the uncertainty of knowledge employing a fuzzy belief function. It describes the learning set with the most appropriate fuzzy if-then rules. For this purpose, a new method of evaluation of membership functions in a set of rules and their iterative refinement is proposed. Results: EasIeR achieves the best results of trade-off criteria between the generalization quality and the interpretability of the knowledge base when compared with state-of-art fuzzy rule-based classification approaches. Achieving the simplest rule base and competitive generalization quality is verified by cross-validation experiments with benchmark data sets. Significance: The new method of rule evaluation allows for better determination of their quality in decision support, and the selection algorithm quickly provides a clear and easy to interpret knowledge base. This rule evaluation technique will be proficient for all fuzzy rule-based classification systems.