Learning Uncertainty with Artificial Neural Networks for Improved Predictive Process Monitoring

TL;DR

This paper enhances predictive process monitoring by integrating Bayesian neural networks to estimate uncertainty, enabling more accurate, confident, and early predictions in process management tasks.

Contribution

It introduces methods to incorporate model and observational uncertainty into neural networks for process monitoring, improving prediction confidence and early-stage performance.

Findings

Uncertainty estimates differentiate prediction accuracy levels.

Confidence intervals can be constructed for predictions.

Techniques are fast and improve early-stage predictions.

Abstract

The inability of artificial neural networks to assess the uncertainty of their predictions is an impediment to their widespread use. We distinguish two types of learnable uncertainty: model uncertainty due to a lack of training data and noise-induced observational uncertainty. Bayesian neural networks use solid mathematical foundations to learn the model uncertainties of their predictions. The observational uncertainty can be calculated by adding one layer to these networks and augmenting their loss functions. Our contribution is to apply these uncertainty concepts to predictive process monitoring tasks to train uncertainty-based models to predict the remaining time and outcomes. Our experiments show that uncertainty estimates allow more and less accurate predictions to be differentiated and confidence intervals to be constructed in both regression and classification tasks. These conclusions remain true even in early stages of running processes. Moreover, the deployed techniques are fast and produce more accurate predictions. The learned uncertainty could increase users' confidence in their process prediction systems, promote better cooperation between humans and these systems, and enable earlier implementations with smaller datasets.