Misclassification-Error-Inspired Ensemble of Interpretable First-Order TSK Fuzzy Subclassifiers: A Novel Multiview Learning Perspective

Abstract

This study explores a novel interpretable Takagi–Sugeno–Kang (TSK) fuzzy ensemble classifier called MEI-TSK from a multiview learning perspective. Unlike most existing fuzzy ensemble classifiers where aggregation learning performs only after the training of multiple fuzzy subclassifiers, MEI-TSK first allows the determination of the antecedents of all fuzzy rules in an individual way for each of TSK fuzzy subclassifiers. It then employs the proposed misclassification-error-inspired learning to accomplish its ensemble learning by training the consequents of all fuzzy rules of each TSK fuzzy subclassifier in a multiview learning way with the simplest averaging aggregation. As a result, the misclassification error caused by such an ensemble learning of fuzzy subclassifiers is theoretically upper bounded. MEI-TSK also features the use of both Bernoulli random feature selection and random feature permutation. The permuted features can be conveniently useful for determining all the antecedents of fuzzy rules with diversity guarantee among all the subclassifiers, and accordingly, be discarded after ensemble learning, resulting in shorter fuzzy rules and improved generalization capability. The experimental results indicate the effectiveness of MEI-TSK in terms of classification performance and/or interpretability.