Comparative Evaluation of Post-Hoc Explainability Methods in AI: LIME, SHAP, and Grad-CAM

This research study presents a comparative evaluation of three prominent post-hoc explain ability methods: LIME (Local Interpretable Model-agnostic Explanations), SHAP (SHapley Additive exPlanations), and Grad-CAM (Gradient-weighted Class Activation Mapping). The evaluation is performed using a simple Convolutional Neural Network (CNN) model trained on the MNIST dataset. To achieve high accuracy, the CNN model was optimized through rigorous training and evaluation, reaching a test accuracy of approximately 98%. The evaluation involves applying each method to selected test images and comparing their effectiveness in terms of interpretability and visualization. LIME provides local feature importance by highlighting influential regions in the images, while SHAP offers a model-agnostic perspective with Shapley values that quantify feature contributions. Grad-CAM visualizes class-specific activation maps, showing which parts of the image contribute to predictions This study provides valuable insights into the strengths and limitations of each explainability method, helping practitioners understand their practical applications and choose the most suitable technique based on their specific needs.