Riess 2024 — SynBPS: Parametric Simulation Framework for Event-Log Data

Single-authored article in SIMULATION (SCS). Corresponds to Paper II of Riess's PhD thesis. Authored during his transition from NMBU to Telenor Research.

Summary

The paper opens with a methodological critique: virtually all published PPM models are evaluated on a small set of public event logs (BPIC, Sepsis, Helpdesk, etc.). Each event log represents one process from one organisation, so a common PPM evaluation benchmark is effectively sampling size N=3–9 process variants — and the research-question logic is almost always to prove that a new method beats a baseline on these specific logs. When external validity is the goal, this is insufficient. Worse, when the researcher wants to understand which data-generating-process characteristics drive performance differences (e.g., Tax et al. 2017 attributing a bad result on one log to repeating activities), only qualitative post-hoc coding is possible; control variables and interaction testing are not.

Two research questions:

RQ1 — to what extent do current BPS frameworks support model-robustness assessment in PPM?
RQ2 — how can their limitations be overcome by a new framework?

RQ1 is answered via a targeted review of open-source BPS tools (Camargo et al. Simod; López-Pintado & Dumas; Pourbafrani et al. SimPT/PMSD; Fracca et al. bpsimpy; Peeperkorn et al.; Grüger et al. SAMPLe; Arena; CPN Tools). Common denominator: these tools optimise for ecological validity (realistic event logs for what-if analysis of specific processes) — their control flows are learned or hand-specified from observed behaviour. They are not designed for the opposite problem: generating synthetic event logs from theoretical data-generating processes with independent manipulation of complexity, memory, stability.

RQ2 introduces SynBPS. Design choices:

Control flow = discrete-time Markov chain, order k configurable (k=1 memoryless, higher-order with D × D^k transition tensor). Absorbing state terminates the trace.
Entropy control over the transition matrix: minimum (deterministic), maximum (uniform/random), medium (n transitions across state space of size d).
Arrivals = Poisson process.
Activity durations = exponential / hypoexponential distributions (chosen for analytic tractability of trace durations).
Stability: process parameters can be drifted deterministically or stochastically to generate concept-drift-exposed logs.
Output = standard event-log tables (case ID, activity, timestamp, resource, attributes).

The paper demonstrates generation of 2304 unique synthetic event logs across manipulated parameters; three representative control flows (minimum, medium, maximum entropy) are inspected via inductive miner + PM4Py.

Acknowledged limitations: Markov chains do not natively model parallel/concurrent activities (requires state-space explosion); resource behaviour and interruptions cannot be modelled explicitly (they are subsumed into duration distributions), which makes SynBPS less suitable for prescriptive-process-monitoring research; duration distributions are restricted to exponential/hypoexponential in the initial release.

Future directions: conditional transition matrices (attribute-dependent routing), time-dependent rate parameters for duration distributions, mixture distributions.

Connections

Paper II of sources/2023-riess-phd-thesis-ppm.
Establishes concepts/business-process-simulation as a PPM-evaluation methodology — complementing, not replacing, calibrated BPS tools.
Explicitly contrasts with BPS tools reviewed: Camargo et al. Simod (see also sources/2021-dumas-process-mining-2-from-insights-to-action), López-Pintado & Dumas, Pourbafrani et al., Chapela-Campa & Dumas family.
Cites sources/2017-tax-lstm-process-prediction, sources/2017-navarin-lstm-data-aware-remaining-time, sources/2016-teinemaa-structured-unstructured-ppm, sources/2019-verenich-survey-ppm, sources/2011-vanderaalst-process-mining-book.
Concepts: concepts/business-process-simulation, concepts/predictive-process-monitoring, concepts/behavioral-variability.
Methods: methods/process-simulation, methods/process-mining-basics.
Author: entities/mike-riess; affiliation entities/telenor (primary) + entities/nmbu (secondary).