A Perspective on Explainable Artificial Intelligence Methods: SHAP and LIME

Explainability tools like SHAP and LIME don't provide universal answers: Both methods are "highly affected by the adopted ML model and feature collinearity," meaning their explanations can change significantly depending on which machine learning model is used and how correlated the input features are. This suggests users shouldn't treat their outputs as objective truth.

These tools are most reliable for specific data types: SHAP and LIME are "particularly" effective with tabular data (structured data in rows and columns), so their reliability may be lower for other data formats like images or text, which is important to know before deploying them in different contexts.

Feature relationships matter more than we might think: The paper highlights that collinearity (when input features are highly correlated with each other) significantly impacts how these explainability methods work, suggesting that understanding your data's structure is just as important as understanding the model itself.

Explainability requires careful interpretation, not blind trust: While these methods aim to "increase the trust of end-users into their output," the paper's cautionary findings suggest that trust should be informed and critical—users need training on how to properly interpret these explanations rather than accepting them at face value.