Residual Sketch Learning for a Feature-Importance-Based and Linguistically Interpretable Ensemble Classifier.

Area of Science:

• * Computational Intelligence
• * Machine Learning
• * Fuzzy Systems

Background:

• * Traditional classifiers often lack interpretability or require complex structures.
• * Linear regression models are recognized for their simplicity and interpretability.
• * Cognitive behavior progresses from coarse to fine analysis.

Purpose of the Study:

• * To propose a novel hybrid ensemble classifier, the hybrid Takagi-Sugeno-Kang fuzzy classifier (H-TSK-FC).
• * To introduce a residual sketch learning (RSL) method for enhanced classifier performance and interpretability.
• * To achieve both feature-importance-based and linguistic-based interpretability in a single model.

Main Methods:

• * Developed H-TSK-FC, integrating deep/wide fuzzy classifier virtues with linear regression components.
• * Implemented RSL: a global linear regression for feature importance, residual partitioning, and parallel TSK fuzzy subclassifiers using ESSC and LLM.
• * Employed minimal-distance-based priority for final predictions to boost generalization.

Main Results:

• * H-TSK-FC demonstrates faster running speeds compared to existing interpretable TSK fuzzy classifiers.
• * Achieved superior linguistic interpretability, characterized by fewer rules and reduced model complexity.
• * Maintained at least comparable generalization capability despite improvements in speed and interpretability.

Conclusions:

• * H-TSK-FC offers a compelling alternative for interpretable classification tasks.
• * The RSL method effectively refines local predictions and enhances overall model performance.
• * Feature-importance-based interpretability contributes to improved efficiency and understandability in fuzzy classification.