TSK Fuzzy Model with Integrated Label Smoothing and Feature Smoothing

The TSK fuzzy model, proposed by Takagi, Sugeno, and Kang, transforms complex nonlinear problems into linear problems on different small segments and uses multiple linear subsystems to fit the nonlinear system. When the TSK fuzzy model is applied to imbalanced data, its learning performance is easily affected by data imbalance, resulting in poor generalization ability. Real-world data often exhibits imbalanced distributions, and existing techniques for handling imbalanced data focus on independent classification objectives. However, many tasks involve classification indicators without obvious hard boundaries. This paper introduces two methods (label smoothing and feature smoothing) to handle the imbalanced classification of labels using the similarity of neighboring targets in the label space and feature space. The specific weighting method and the fusion of the TSK model are used to improve the generalization ability of the TSK fuzzy model and ensure the interpretability of the obtained rules. Experimental results show that the integrated TSK fuzzy model has good performance for imbalanced classification.