A multi-view TSK fuzzy system with deformable Gaussian membership functions and rule-level attention for...

Area of Science:

• Computational Intelligence
• Machine Learning
• Fuzzy Systems

Background:

• Traditional fuzzy systems often lack flexibility in high-dimensional, complex classification tasks.
• Existing models struggle with dynamic adaptation to varying data characteristics and rule importance.
• Multi-view learning presents challenges in integrating heterogeneous data sources effectively.

Purpose of the Study:

• To develop a novel multi-view Takagi-Sugeno-Kang (TSK) fuzzy system (MDA-TSK-FS) with enhanced modeling capacity and flexibility.
• To improve the adaptability of fuzzy rules to input space variations using deformable membership functions.
• To dynamically optimize rule importance and selection through a rule-level attention mechanism.

Main Methods:

• Integration of deformable Gaussian membership functions with learnable deformation offsets in the antecedent part.
• Implementation of a multi-head attention mechanism at the rule level for adaptive weight allocation based on sample-specific information.
• Extensive experimentation on five public multi-view datasets: Caltech7, Handwritten, Dermatology, Forest, and EEG.

Main Results:

• The proposed MDA-TSK-FS achieved superior classification accuracies: 94.38% (Caltech7), 98.62% (Handwritten), 98.58% (Dermatology), 88.57% (Forest), and 69.75% (EEG).
• Ablation studies confirmed the significant contributions of the deformable antecedent structure (approx. 7% EEG accuracy increase) and rule-level attention (approx. 6% EEG accuracy increase).
• The model demonstrated improved generalization and interpretability, especially for multi-source heterogeneous data, outperforming strong baselines.

Conclusions:

• The proposed MDA-TSK-FS offers a novel and effective modeling paradigm for multi-view fuzzy inference systems.
• Deformable membership functions and rule-level attention mechanisms significantly enhance fuzzy system performance and adaptability.
• The approach holds substantial theoretical value and practical potential for complex, high-dimensional classification tasks involving multi-view data.