Machine Learning Model Interpretability and Explainability

Abstract

The increasing deployment of complex machine learning models in high-stakes domains has created an urgent need for interpretability and explainability. This study investigates the effectiveness of various explainable artificial intelligence (XAI) techniques across three critical dimensions: faithfulness, stability, and comprehensibility. Using a mixed-methods design combining quantitative evaluation with user-centric assessment, we examined five XAI methods—LIME, SHAP, Integrated Gradients, Layer-wise Relevance Propagation (LRP), and Attention Mechanism Visualization (AMV)—across multiple model architectures and two downstream tasks. Quantitative results demonstrate that LIME consistently achieves the highest Human-reasoning Agreement (MAP = 0.82), while AMV shows superior robustness with a Mean Average Difference of 0.24 against input perturbations. LRP excels in contrastivity, particularly with complex transformer models. However, user studies with 39 participants across three expertise groups revealed that technical accuracy of explanations does not guarantee user comprehension, with non-experts understanding only 58% of explanation content compared to 89% for experts. Qualitative findings indicate that domain knowledge integration significantly improves explanation quality, with semantically structured representations improving alignment with expert reasoning by 43%. These findings contribute to a nuanced understanding of XAI effectiveness and provide practical guidelines for designing interpretable machine learning systems that balance technical fidelity with human comprehensibility.