FULL PDF
Abstract
The rapid adoption of machine learning models in high-stakes domains demands transparency and accountability. However, complex nonlinear predictive models, such as deep neural networks and ensemble methods, are often perceived as "black boxes," limiting user trust and model adoption. This paper systematically compares prominent post hoc interpretability techniques available, including LIME, SHAP, and partial dependence plots (PDPs), evaluating their strengths, limitations, and suitability across different types of complex models. Our comparative study identifies key trade-offs between local and global interpretability, computational overhead, and faithfulness of explanations. Through both theoretical analysis and empirical evaluation, we aim to guide practitioners in selecting appropriate interpretability methods based on their use cases.
View more >>