Mostafa Abbasi

Topic

From Understanding to Action: A Unified AI-Driven Framework for Predictive, Representational, and Prescriptive Business Process Management

Department of Mechanical Engineering

Date & location

• Wednesday, October 7, 2026

• 2:00 P.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Homayoun Najjaran, Department of Mechanical Engineering, University of Victoria (Supervisor)

• Dr. Keivan Ahmadi, Department of Mechanical Engineering, UVic (Member)

• Dr. Patricia Lasserre, Department of Computer Science, UBC (Outside Member) 

External Examiner

• Dr. Babak Mohamadpour Tosarkani, School of Engineering, University of British Columbia (Okanagan)

Chair of Oral Examination

• Dr. Falk Herwig, Department of Physics and Astronomy, UVic

   

Abstract

Business Process Management (BPM) concerns the analysis, improvement, and automation of organizational workflows. Artificial intelligence and machine learning have shifted BPM from reactive control toward proactive, evidence-based decision making, yet despite the widespread use of AI, four gaps persist: i) predictive models often lack uncertainty estimates; ii) conformance methods remain fragile under noisy event logs; iii) discovering coherent process models from partial observations remains an open problem; and iv) prescriptive guidance is confined to binary actions that ignore continuous performance indicators. 

This dissertation addresses these gaps through four connected contributions that form a single foundation for moving from analytical understanding to actionable intelligence. All four start from the same minimal event log information: which case an event belongs to, which activity occurred, and when it occurred. First, an uncertainty-aware framework reconstructs the full activity-level timeline of a running case through a three-stage Quantile XGBoost (QXGBoost) architecture, producing calibrated multi-quantile prediction intervals; across nine real-world logs, it attains 80% interval coverage within ± 5% of nominal on seven datasets and intervals up to 29× tighter than Laplace-based LSTM baselines. Second, a model-centric conformance framework turns Petri nets into graphs, learns node embeddings through adaptive directed random walks, and aligns them via Procrustes analysis, yielding an entropy-regularized distance that resists log noise and localizes activity-level deviations. Third, this representational machinery turns generative: a student-teacher transformer distills skip-gram structural priors into a causal decoder that assembles coherent models from overlapping trace subgraphs; mining its traces outperforms mining the original log on eight of nine logs (mean f-score 0.77 vs. 0.44), while a lightweight construction reaches 0.9046 ± 0.0517, a 19.99% gain over the strongest generative baseline. Fourth, FORLAPS (Fine-Tuned Offline Reinforcement Learning Augmented Process Sequence Optimization) couples offline Q-learning with state dependent reward shaping and process-aware augmentation; across nine logs spanning healthcare, finance, public administration, and industry, it produces recommendations closest to desired-outcome traces on eight logs while preserving control-flow feasibility by construction, rather than collapsing onto the behaviour policy.