Prof. Marlon Dumas (Institute of Computer Science UT),
Assoc. Prof. Fabrizio Maggi (Institute of Computer Science, UT),
Prof. Marcello La Rosa (University of Melbourne, Australia),
Prof. Arthur ter Hofstede (Queensland University of Technology, Australia)
Prof. Peter Fettke (German Research Center for Artificial Intelligence),
Prof. Alfredo Cuzzocrea (University of Trieste and ICAR-CNR, Italy).
Modern enterprise systems collect detailed data about the execution of the business processes they support. The widespread availability of such data in companies, coupled with advances in machine learning, have led to the emergence of data-driven and predictive approaches to monitor the performance of business processes. By using such predictive process monitoring approaches, potential performance issues can be anticipated and proactively mitigated. Various approaches have been proposed to address typical predictive process monitoring questions, such as what is the most likely continuation of an ongoing process instance, or when it will finish. However, most existing approaches prioritize accuracy over explainability. Yet in practice, explainability is a critical property of predictive methods. It is not enough to accurately predict that a running process instance will end up in an undesired outcome. It is also important for users to understand why this prediction is made and what can be done to prevent this undesired outcome.
This thesis proposes two methods to build predictive models to monitor business processes in an explainable manner. This is achieved by decomposing a prediction into its elementary components. For example, to explain that the remaining execution time of a process execution is predicted to be 20 hours, we decompose this prediction into the predicted execution time of each activity that has not yet been executed. We evaluate the proposed methods against each other and various state-of-the-art baselines using a range of business processes from multiple domains. The evaluation reaffirms a fundamental trade-off between explainability and accuracy of predictions.
The research contributions of the thesis have been consolidated into an open-source tool for predictive business process monitoring, namely Nirdizati. It can be used to train predictive models using the methods described in this thesis, as well as third-party methods. These models are then used to make predictions for ongoing process instances; thus, the tool can also support