A Comparative Study of SHAP and LIME: Interpreting Black-Box Machine Learning Models

Explainable Artificial Intelligence (XAI) has become very important for explaining the decision-making processes of complex machine learning (ML) and deep learning (DL) models. As the grouping methods and deep neural networks demonstrate high predictive performance, as the nature of black-box it increases the difficulty for adoption in sensitive domains such as healthcare and finance. So, two most popular post-hoc techniques which are used for explanation are compared in this study, (1) SHAP- Shapley Additive Explanation (2) LIME- Local Interpretable Model-Agnostic. To assess this two we are using two-three publicly available datasets and also multiple ML models. Assessing them by evaluating key aspects like computational efficiency, consistency, importance, and fidelity. Comparing them gives us that SHAP offers higher stability and helps alignment in domain, while on the other hand LIME provides faster albeit more variable and insights. All the comparison made in this study aims to guide practitioners in choosing the appropriate method to interpret.