Analysis, Comparison and Interconnection of Business Process Suffix Prediction and Marked-Temporal Point Processes

Provided and advised by: Henryk Mustroph

Business process suffix prediction aims to forecast the remaining behavior of an ongoing process instance based on the events observed so far. Given a partial trace, or prefix, such models predict the sequence of future events, such as future activities and often their associated timestamps until process completion. This line of research originates from predictive process monitoring [1] and has become increasingly relevant for operational decision support, early risk detection, and what-if analysis in complex organizational processes. Accurate suffix prediction is particularly challenging due to concurrency, loops, data-dependent decisions, and uncertainty in both control flow and timing. The most recent suffix prediction approach is probabilistic suffix prediction [2]. Unlike deterministic approaches that output a single most likely continuation, PSP models generate multiple plausible future traces, typically by sampling from an autoregressive sequence model (e.g., an encoder–decoder LSTM or a Transformer).

 Marked Temporal Point Processes (MTPPs) [3] provide a mathematically grounded framework for modeling sequences of events in continuous time, where each event is associated with a discrete mark (e.g., an activity label) and a stochastic intensity function governing event occurrence. In recent years, neural MTPP models, such as recurrent, attention-based, or Transformer-based intensity models, have shown strong performance in domains like healthcare, user behavior modeling, and language-like event streams. These models naturally capture temporal dynamics and uncertainty, offering a principled probabilistic alternative to many sequence-to-sequence approaches commonly used in business process suffix prediction.

 This thesis is motivated by the observation that business process suffix prediction and MTPPs address closely related problems from different perspectives, yet their relationship is not well understood in the process mining literature. By analyzing and comparing these paradigms, and by studying how they can be interconnected or unified, this work aims to clarify modeling assumptions, strengths, and limitations on both sides.

Recommended Readings:
[1] Di Francescomarino, C., & Ghidini, C. (2022). Predictive process monitoring. In Process mining handbook (pp. 320-346). Cham: Springer International Publishing.
[2] Mustroph, H., Kunkler, M., & Rinderle-Ma, S. (2025). An Uncertainty-Aware ED-LSTM for Probabilistic Suffix Prediction. arXiv preprint arXiv:2505.21339.
[3] Shchur, O., Türkmen, A. C., Januschowski, T., & Günnemann, S. (2021). Neural temporal point processes: A review. arXiv preprint arXiv:2104.03528.

The application must contain:
- Thesis topic you are interested in.
- Current Transcript of Records.
- Application form
- CV
Please send the application to master.i17(at)in.tum.de AND (in cc) henryk.mustroph(at)tum.de.