Local Surrogate Models With Residual Fuzzy Rules for Model-Agnostic Explanations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI (XAI)

Background:

• Complex black-box regression models lack transparency.
• Local Interpretable Model-Agnostic Explanations (LIME) provide local fidelity but are sensitive to parameters.
• Linear regression is commonly used in LIME for its simplicity and interpretability.

Purpose of the Study:

• To design a linear regression model for local explanations of black-box models.
• To enhance the local fidelity of linear explanations using fuzzy residual rules.
• To improve the precision and conciseness of AI model explanations.

Main Methods:

• Developed a model-agnostic system with three components: optimal kernel size strategy, a local linear regression model, and fuzzy rules for residuals.
• Constructed fuzzy rules based on errors from the local linear regression model.
• Implemented Lasso regression for feature selection to enhance local model interpretability.

Main Results:

• The proposed architecture achieved higher precision in explanations.
• Interpretable fuzzy rules provided a more concise description of black-box model characteristics.
• The system effectively enhances the accuracy and understandability of local AI explanations.

Conclusions:

• The novel approach successfully improves the fidelity and interpretability of local explanations for black-box regression models.
• Fuzzy residual rules offer a valuable mechanism for refining LIME-based explanations.
• The integration of feature selection further boosts the interpretability of local models.