Fuzzy classification involved in fusion of existing decision and pre-known task applied for integrated input space

Abstract

Although the Takagi-Sugeno-Kang (TSK) fuzzy classifier has achieved great success, how to further improve its classification performance and enhance its interpretability is still one of the most difficult challenges. Involved with the fusion of existing decision information and pre-known classification task, a newly proposed deep/hierarchical TSK fuzzy classifier (EDIPK-TSK) with interpretable fuzzy rules makes full use of the classification advantages of each base classifier to construct a multi-layer deep learning structure. This study first considers that the existing decision information of each training sub-block is sequentially projected into the subsequent sub-blocks for training. Undoubtedly, the existing decision information has played a guiding role in the current learning process to some extent. Simultaneously, the pre-known classification task is fused into the decision information for fine-tuning of it, which can significantly improve the efficiency of guidance and accelerate the fitting speed of the model. In each layer, the use of interpretable integration input space guarantees that EDIPK-TSK is not a black box. The proposed deep classifier can realize learning by using short fuzzy rules, which ensures the satisfactory interpretability of the classifier. The final experimental results also verify that EDIPK-TSK has strong classification advantages and interpretability.