ExplainabilityTransparencyTrustworthinessstat.ML
Characterizing the contribution of dependent features in XAI methods
Ahmed Salih, Ilaria Boscolo Galazzo, Zahra Raisi-Estabragh, Steffen E. Petersen, Gloria Menegaz, Petia Radeva
April 4, 2023
IEEE journal of biomedical and health informatics
31 citations
Key Takeaways
-
Popular XAI methods like SHAP and LIME may give misleading results when the features they analyze are correlated with each other, since these methods assume features are independent—a condition rarely true in real-world data.
-
Feature importance rankings can be unreliable when collinearity (correlation between predictors) is present, potentially identifying the wrong variables as most important for model decisions, which undermines trust in AI explanations.
-
A simple adjustment method can improve XAI reliability by accounting for feature dependencies, allowing the approach to work with any existing explanation method without requiring changes to the underlying model.
-
This fix is practical and accessible because it is model-agnostic (works across different AI systems) and straightforward to calculate, making it feasible for organizations to implement when evaluating their AI systems' explanations.
Abstract
Explainable Artificial Intelligence (XAI) provides tools to help understanding how the machine learning models work and reach a specific outcome. It helps to increase the interpretability of models and makes the models more trustworthy and transparent. In this context, many XAI methods were proposed being SHAP and LIME the most popular. However, the proposed methods assume that used predictors in the machine learning models are independent which in general is not necessarily true. Such assumption casts shadows on the robustness of the XAI outcomes such as the list of informative predictors. Here, we propose a simple, yet useful proxy that modifies the outcome of any XAI feature ranking method allowing to account for the dependency among the predictors. The proposed approach has the advantage of being model-agnostic as well as simple to calculate the impact of each predictor in the model in presence of collinearity.