Coalitional Strategies for Efficient Individual Prediction Explanation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning (ML) models are increasingly used in research and industry.
• Understanding ML model predictions, especially for non-experts, is crucial but challenging due to the
• black box
• nature of these models.
• Existing explanation methods often require long computation times or rely on restrictive assumptions, failing to fully capture attribute interactions.

Purpose of the Study:

• To introduce and evaluate new methods for explaining ML model predictions.
• To identify relevant groups of attributes, termed "coalitions," that influence predictions.
• To compare the efficiency and accuracy of these coalition-based methods against existing literature, including SHapley Additive exPlanation (SHAP).

Main Methods:

• Developing and implementing methods based on the detection of attribute coalitions.
• Comparing the performance of coalition methods with established techniques like SHAP.
• Assessing computation time and the accuracy of individual prediction explanations.

Main Results:

• Coalition-based methods demonstrate greater efficiency compared to existing approaches like SHAP.
• These methods significantly reduce computation time while maintaining acceptable accuracy for individual prediction explanations.
• The findings highlight the effectiveness of attribute coalitions in providing interpretable ML insights.

Conclusions:

• Coalition methods offer a more efficient and practical approach to explaining ML model predictions.
• The improved efficiency enables wider adoption of explanation techniques, fostering greater trust between ML models, users, and stakeholders.
• This work contributes to making complex ML systems more transparent and understandable.