Comprehensive Framework for Sensitivity Analysis in Decision Tree-Based Classification

This paper presents a framework for performing the sensitivity analysis of a Decision Tree-based classification method without relying on iterative techniques such as Monte Carlo simulations. A preliminary model is introduced to identify the main sources of uncertainty in Machine Learning systems used for classification or regression tasks. Based on this model, a statistical formulation is developed to propagate the marginal probability density functions of the input measurements through a trained Decision Tree classifier, enabling the estimation of the probability of correct classification for a single instance. Sensitivity analysis results obtained with the proposed approach are presented and compared with Monte Carlo-based estimates, showing agreement and thus demonstrating the effectiveness of the method.