Doctoral defence: Mahmoud Shoush "Prescriptive Process Monitoring Under Uncertainty and Resource Constraints“

On 29 May 2025 at 14:15 Mahmoud Kamel Akila Soliman Shoush will defend his doctoral "Prescriptive Process Monitoring Under Uncertainty and Resource Constraints“ to obtain the degree of Doctor of Philosophy (in Computer Science).

Supervisor
Prof. Marlon Dumas, University of Tartu

Opponents
Prof. Annalisa Appice, University of Bari Aldo Moro (Italy)
Assoc. Prof. Sylvio Barbon Junior, University of Trieste (Italy)

Summary
A university admissions team reviews applicants' documents, collaborates with committees to decide on acceptances, and communicates results—steps in the student admission business process. While some applications proceed smoothly, others face delays due to missing documents or errors. Similar issues occur in business processes like bank loan approvals or online product returns.

Advances in machine learning can predict issues, such as missing documents or product returns, before they occur. Yet, these predictions are only useful if followed by actions, called "interventions," to prevent undesired outcomes. For example, if a system predicts a loan applicant will reject an offer, it might recommend an intervention, such as sending a personalized offer, to increase acceptance chances.

In this example, a simple approach might trigger interventions when the rejection probability exceeds 80%, but this has three limitations: (1) The intervention may be ineffective if the applicant rejects the loan due to factors beyond the bank's control, such as high interest. (2) Predictions come with uncertainty, and it may be better to ignore highly uncertain ones. (3) Sending personalized offers requires time and resources, limiting how many can be sent daily.

In this PhD thesis, we argue that a method to turn predictions into interventions needs to consider the following factors: Is the intervention needed? Will it be effective? Is it urgent, or can it be delayed? Do we have the capacity to perform it? And how uncertain are we about its need, effectiveness, and urgency?

We propose three machine learning-based methods for triggering interventions: the first addresses prediction uncertainty, the second considers all factors, allowing users to set rules, and the third automatically learns the best intervention strategy without predefined rules. Using real-world data, we show that these methods allow more effective, uncertainty- and resource-aware interventions, optimizing business processes.

The defence will be held also in Zoom (meeting ID: 670 504 9543, passcode: ati).