Characterizing the Contribution of Dependent Features in XAI Methods.

Area of Science:

• Artificial Intelligence
• Machine Learning Interpretability
• Medical Data Analysis

Background:

• Explainable Artificial Intelligence (XAI) methods enhance model transparency but struggle with multicollinearity.
• Existing XAI techniques often assume feature independence, which is not always valid.
• Multicollinearity can compromise the reliability of feature importance rankings in XAI.

Purpose of the Study:

• To propose a novel proxy to address multicollinearity in XAI feature ranking.
• To enhance the robustness and interpretability of XAI outcomes.
• To provide a method that accounts for feature dependencies.

Main Methods:

• A simple proxy was developed to modify existing XAI feature ranking outcomes.
• The proxy was applied to the SHAP (SHapley Additive exPlanations) method.
• Cardiac magnetic resonance imaging data was used for a male vs. female classification task with nine cardiac phenotypes as features.
• Principal Component Analysis and biological plausibility were used for validation.

Main Results:

• The proposed proxy effectively accounts for feature dependency in XAI.
• The modified feature rankings were more robust in the presence of multicollinearity compared to original SHAP.
• The method successfully revealed the impact of dependent features on model outcomes.

Conclusions:

• The developed proxy offers a valuable improvement for XAI methods dealing with multicollinearity.
• This approach leads to more reliable identification of informative features.
• The findings contribute to more trustworthy and transparent AI in complex datasets.