A New Metric Based on Association Rules to Assess Feature-Attribution Explainability Techniques for Time Series...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Explainable AI (XAI) methods are crucial for understanding model decisions.
• Current XAI evaluation lacks a unified, model-independent metric.
• Quantifying the quality of feature attributions remains a challenge.

Purpose of the Study:

• Introduce RExQUAL, a novel, model-independent metric for evaluating attribution-based XAI techniques.
• Provide a quantitative framework to compare the quality of explanations.
• Integrate local and global explanation insights.

Main Methods:

• Utilize feature attribution from model-agnostic XAI methods.
• Generate association rules using key attributes for forecasting tasks.
• Propose global support and confidence metrics for rule quality assessment.
• Combine association rule metrics for a comprehensive explanation quality score.

Main Results:

• RExQUAL effectively quantifies explanation quality.
• The metric demonstrates versatility across different time series forecasting tasks.
• Comparative analysis shows RExQUAL's efficacy in evaluating SHAP, LIME, and RULEx.

Conclusions:

• RExQUAL offers a robust, quantitative framework for XAI evaluation.
• The proposed metric facilitates reliable comparison of diverse explainability techniques.
• This work advances the field of XAI by providing a standardized quality assessment tool.