Enhancing TSK Fuzzy Systems with an Automated Filter-Gate Mechanism for Rule-Specific Feature Selection
Yingtao Ren, Yu‐Cheng Chang, Xiaowei Jiang, Jie Yang, Chin‐Teng Lin · 2025
The Takagi-Sugeno-Kang (TSK) fuzzy system has been extensively applied across a wide range of real-world scenarios. However, in most rule-based TSK fuzzy systems, all input features are naively included in each fuzzy rule, leading to additional memory and computational costs. Even worse, redundant membership functions from some features may introduce noise and complicate the training process. To address these issues, we propose a two-stage algorithm called Automated Filter-Gate (AutoFG) that automatically filters redundant features for each rule and can be applied to different TSK fuzzy systems. In the search stage, we relax the selection by introducing learnable gate parameters with Regularization rather than relying on expert prior knowledge or exhaustive searching methods. Through this approach, AutoFG learns sparse gate parameters through gradient descent, automatically identifying and removing redundant features (membership functions) in each rule. In the re-train stage, we remove low-contribution features from the corresponding rules, restore the membership functions of retained features, and fine-tune the model to maintain stable performance. Experiments on twelve classification datasets demonstrate that AutoFG effectively filters out redundant features in a rule-specific manner. Moreover, TSK fuzzy systems that incorporate AutoFG achieve a substantial performance improvement.