Style-constrained Takagi-Sugeno-Kang Fuzzy Classifier

Abstract

Due to the excellent interpretability and classification performance, the Takagi-Sugeno-Kang fuzzy classifier (TSK-FC) has drawn great attention. However, different patterns own their respective homogeneities, and the samples from different groups present explicitly or implicitly homogenous styles, which are significantly different from the assumption that all samples from different groups are identically and independently distributed (i.i.d.). In this paper, a style-constrained Takagi-Sugeno-Kang fuzzy classifier called SC-TSK-FC is proposed by breaking the i.i.d. assumption. To explore the styles of data, a series of style matrices are embedded into the objective function of the TSK-FC. Besides, with the introduction of the regularization term corresponding to each style matrix, the nuances between different styles of data can be captured, which can help improve the classification performance of SC-TSK-FC. Particularly, five fixed fuzzy partitions with interpretable linguistic terms are adopted to along each input feature, and SC-TSK-FC outperforms TSKFC by using less fuzzy rules. These guarantee the interpretability of SC-TSK-FC. Experimental results on some benchmark datasets demonstrate the effectiveness of the proposed SC-TSK-FC.