A STUDY OF EXPLAINABLE ARTIFICIAL INTELLIGENCE FOR INTERPRETABLE MACHINE LEARNING MODELS

Abstract

This study investigates Explainable Artificial Intelligence (XAI) frameworks for improving the interpretability of machine learning (ML) models in high-stakes decision-making environments. Despite strong predictive performance, many ML models operate as “black boxes,” limiting transparency, accountability, and user trust in domains such as healthcare, finance, and education. The study is grounded in interpretability theory and cognitive trust frameworks, emphasizing post-hoc explainability techniques including Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). A quantitative experimental design was employed using benchmark datasets from UCI and Kaggle repositories. Multiple ML models, including Random Forest, Support Vector Machines, and Gradient Boosting classifiers, were evaluated. XAI techniques (LIME and SHAP) were applied to assess global and local interpretability. Performance was measured using accuracy, F1-score, explanation fidelity, and interpretability indices.Findings indicate that ensemble models achieved higher predictive accuracy, while SHAP-based explanations provided more consistent feature attribution compared to LIME. A minimal reduction in accuracy (1–3%) was observed when interpretability constraints were introduced, indicating an acceptable trade-off between performance and transparency. The integration of XAI improved model interpretability scores by up to 35%, increased user trust indicators in evaluation settings, and maintained over 90% predictive performance across datasets.