Interpreting Machine Learning Predictions with SHAP and LIME for Transparent Decision Making

Abstract

Machine learning models increasingly influence critical decisions across diverse domains, yet their complex architectures often operate as black boxes, obscuring the rationale behind predictions and limiting stakeholder trust. This research demonstrates a comprehensive, reproducible workflow for applying explainable artificial intelligence techniques to interpret Random Forest classifier decisions using publicly available data and standard computational resources. The study implements and compares two leading explanation methods, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), on the Titanic survival prediction task to evaluate their consistency and practical utility. The methodology encompasses automated data preprocessing, model training with regularization to prevent overfitting, and systematic generation of both global and local explanations through multiple visualization formats. Results reveal exceptional agreement between explanation methods, with Spearman rank correlation of 0.918 and Pearson correlation of 0.982 for feature importance values. Both techniques consistently identified passenger sex as the dominant predictive feature, contributing approximately 15.4% and 12.5% of model decisions respectively, followed by passenger class and fare. The Random Forest model achieved 84.5% test set ROC-AUC with controlled overfitting (0.040 ROC-AUC gap between training and test sets) while maintaining interpretable complexity through architectural constraints. The implementation executes efficiently on CPU hardware within minutes, eliminating computational barriers to XAI adoption. This work establishes that current explainability techniques can provide reliable, consistent insights into ensemble model decisions while remaining accessible to researchers and practitioners with limited computational resources.