A comprehensive analysis of perturbation methods in explainable AI feature attribution validation for neural time...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Explainable AI (XAI) is crucial in high-stakes domains like medicine and finance.
• Feature attribution methods (AMs) are common for identifying influential features in AI models.
• Current validation metrics for AMs have shown flaws, particularly with time series data.

Purpose of the Study:

• To address the need for rigorous validation of explanation methods in AI.
• To introduce a novel metric for faithful assessment of feature importance attribution.
• To develop an adapted methodology for robust faithfulness evaluation of AMs.

Main Methods:

• Introduction of the Consistency-Magnitude-Index metric for AM validation.
• Development of an adapted methodology using diverse perturbation methods for faithfulness evaluation.
• Extended evaluation of AMs on time series data, considering perturbation methods and region size.

Main Results:

• The Consistency-Magnitude-Index facilitates a more faithful assessment of feature importance.
• Perturbation methods and region size significantly influence AM evaluation on time series data.
• Guidelines for future AM faithfulness assessments are provided based on extensive evaluation.

Conclusions:

• The proposed metric and methodology enhance the trustworthiness of AI explanations.
• The study provides practical insights for evaluating AMs in time series analysis.
• Demonstration of the methodology on a multivariate time series example validates its utility.