Machine Learning for Predictive and Prescriptive Process Monitoring

Predictive process monitoring (PredPM) is a relatively young and vibrant research field that deals with making predictions about business processes. However, until now, few PredPM methods or applications have found their way into practice. The work presented in this dissertation builds upon and contributes to the current body of literature. At the same time, it aims to pave the road toward practical implementation. For this reason, we focused our efforts on pulling five levers to make progress toward this objective: (1) accuracy, (2) speed, (3) robustness, (4) evaluation and selection, and (5) decision support. The latter lever ``decision support'' surpasses the mere predictions about processes and concerns formulating decisions to actually optimize the outcomes of running processes. Prescriptive process monitoring (PresPM) is the umbrella term for work in this realm.

This dissertation makes valuable contributions to the existing literature in multiple aspects. First, we propose preprocessing techniques to remove bias in the test set and advocate a clean separation of training and test sets to avoid data leakage. These carefully selected preprocessing techniques make models more robust and support the practitioner in evaluating and selecting models. We also argue for the use of benchmarks to make PredPM research results more comparable than they are today.  

Second, we introduce convolutional neural networks (CNN) to both outcome and remaining-time prediction tasks in PredPM. In our experiments, CNN were faster and more accurate than the long short-term memory (LSTM) networks that researchers commonly used in work preceding ours.

Third, we use Bayesian neural networks (BNN) with loss attenuation to train uncertainty-based models to predict the remaining time and outcomes of business processes. In our experiments, uncertainty estimates allowed to evaluate the accuracy of predictions and \textit{select} those exceeding preset accuracy thresholds. Confidence intervals can be constructed based on the uncertainty estimates. Moreover, the deployed techniques were fast and produced more accurate predictions. Learned uncertainty holds the promise to increase users' confidence in their process prediction systems, promote better cooperation between humans and these systems, and enable earlier implementations with smaller datasets.

Fourth, as for decision making support, we conducted experiments on timed process interventions with synthetic data that render genuine online reinforcement learning (RL) and the comparison to causal inference (CI) possible and allow for an accurate evaluation of the results. In our experiments, online RL's policies outperformed those from CI and were more robust at the same time. Nonetheless, CI also produced useful policies and could have its merits in settings where online learning is not an option.

Finally, we describe experiments where the use of uncertainty estimates by an ensemble's stacking model improves its accuracy. We also present a novel neural network cell for PredPM tasks, the multi-dilational convolutional unit (MDCU). In most scenarios in our experiments, MDCU outperformed both CNN and LSTM in terms of accuracy and was much faster than LSTM.

The majority of these techniques and suggestions are ready and relatively easy to implement. By pulling several important levers simultaneously, they should encourage practitioners to make the leap towards the real-life application of PredPM and PresPM for their business processes.