On 5 October at 10:15, Muhammad Awais Ali will defend his doctoral thesis “How Will My Business Process Unfold? Predicting Case Suffixes With Start and End Timestamps”.
Supervisors:
Assoc. Prof. Fredrik Milani, University of Tartu
Prof. Marlon Dumas, University of Tartu
Opponents:
Assoc. Prof. Johannes De Smedt, KU Leuven (Belgium)
Lecturer Vincenzo Pasquadibisceglie, University of Bari Aldo Moro (Italy)
Summary
When a hospital schedules patient appointments, processes insurance claims, or manages staff, it is running a business process, i.e., a structured sequence of tasks carried out to reach a goal. Managers often ask: given a case that is already in progress, what are the remaining steps, and when will they occur? Knowing this information in advance can help organizations to plan workloads, allocate staff, and avoid bottlenecks.
Every task performed in, for instance, a hospital leaves a digital footprint in its IT systems. Modern machine learning can use these historical footprints to learn the business’s "habits," to forecast how an ongoing case will finish. This field is called Predictive Process Monitoring. Specifically, in this field, case suffix prediction is used to generate a complete timeline of all remaining steps until a case is closed.
Current methods, however, have four important blind spots. They confuse waiting time (how long a task sits in a queue) with processing time (how long it takes to do the work). Second, some produce repetitive or unrealistic task sequences. Third, they ignore that many cases compete for the same limited pool of workers, which causes queueing. Therefore, if three cases need the same specialist at the same time, some must wait. Fourth, they overlook real-world constraints such as shifts, working hours, and weekends.
This thesis addresses these four gaps with four contributions: (1) a conceptual framework that distinguishes different types of waiting times; (2) a sampling strategy that avoids repetitive, unrealistic predictions; (3) a prediction method that models competition for shared resources across all ongoing cases; and (4) an extension that incorporates working-time constraints. For instance, the model explicitly reasons that if three cases need the same specialist at once, some must wait, and it predicts which one and by how long.
Evaluated on real-world and simulated data, the proposed approach achieves more accurate predictions of remaining task sequences and lower timestamp errors than state-of-the-art methods, with the greatest gains in resource-constrained settings.
