Augmenting post-hoc explanations for predictive process monitoring with uncertainty quantification via conformalized Monte Carlo dropout

Abstract

This study presents a novel approach to improve the transparency and reliability of deep learning models in predictive process monitoring (PPM) by integrating uncertainty quantification (UQ) and explainable artificial intelligence (XAI) techniques. We introduce the conformalized Monte Carlo dropout method, which combines Monte Carlo dropout for uncertainty estimation with conformal prediction (CP) to generate reliable prediction intervals. Additionally, we enhance post-hoc explanation techniques such as individual conditional expectation (ICE) plots and partial dependence plots (PDP) with uncertainty information, including credible and conformal predictive intervals. Our empirical evaluation in the manufacturing industry demonstrates the effectiveness of these approaches in refining strategic and operational decisions. This research contributes to advancing PPM and machine learning by bridging the gap between model transparency and high-stakes decision-making scenarios.