An Empirical Evaluation of Shapley Additive Explanations: A Military Implication
Trupthi Rao, Sonali Agarwal, Navjot Singh
Abstract
Numerous AI systems have been developed as a result of the progress in machine learning. Currently, there is significant interest in creating state-of-the-art Machine Learning (ML)-based solutions to automate predictions and classification tasks across diverse industries. However, deploying these ML models in sectors that prioritize security and privacy raises concerns about potential bias in predictions. To ensure both accuracy and model comprehensibility, it's crucial to understand how these models operate. In this research, we conducted a comprehensive evaluation of an explain ability/interpretability method known as SHAP using Sensor Information Technology (SensIT) data. Our goal was to shed light on the decision-making process of a black box model classifying two categories of military vehicles. The results revealed that both Tree SHAP and Sampling SHAP provided similar explanations, as assessed in terms of Consistency, Compactness, Stability, and Approximation.