Shapley variable importance cloud for interpretable machine learning.

Area of Science:

• Machine Learning
• Explainable AI
• Statistical Inference

Background:

• Current interpretable machine learning focuses on explaining single, optimized models.
• Shapley Additive Explanations (SHAP) provides local and global model explanations.
• Existing methods may yield biased inference when applied to a single model.

Purpose of the Study:

• To extend global interpretability methods to a set of 'good enough' models.
• To introduce a novel Shapley variable importance cloud for robust model assessment.
• To quantify uncertainty in importance measures for formal statistical inference.

Main Methods:

• Developed Shapley variable importance clouds by pooling Shapley values from multiple relevant models.
• Quantified uncertainty explicitly within the importance measures.
• Created visualizations to highlight uncertainty and its impact on practical inference.

Main Results:

• The Shapley variable importance cloud provides an overall importance measure by integrating information across models.
• Explicit quantification of uncertainty supports formal statistical inference.
• Demonstrated reduced bias in inference compared to single-model SHAP assessments in recidivism and clinical data experiments.

Conclusions:

• The Shapley variable importance cloud offers a more comprehensive and reliable approach to model interpretability.
• This method complements existing SHAP assessments, mitigating potential biases.
• The approach enhances the practical utility of interpretable machine learning in real-world applications.