A design of fuzzy rule-based classifier optimized through softmax function and information entropy

Abstract

Takagi–Sugeno–Kang (TSK) classifiers have achieved great success in many applications due to their interpretability and transparent model reliability for users. At present, however, how to evaluate classification results is still an unsolved issue for TSK classifiers. This study designs a fuzzy rule-based classifier based on TSK classifiers, the outputs of which for an instance can be considered as the membership grades that the instance belongs to all classes. Then, an information entropy-based method is proposed to estimate the certainty of the outputs, which facilitates the further evaluation of the classification results of the instance for users. If the confidence level is not high, users can reject the classification results, and use other more advanced classifiers or collect more information about the instance. Moreover, the developed mechanism is suitable for handling large data since the adaptive moment estimation algorithm is used to identify the parameters of it. Experimental results demonstrate that the developed mechanism outperforms several rule-based classifiers.