Concurrency-Aware Process Model Forecasting with Causal Nets
Yongbo Yu, Jari Peeperkorn, Johannes De Smedt, Jochen De Weerdt
cs.LG · cs.DB
Quick verdict
Extends PMF to concurrency-bearing causal nets; careful evaluation, but concurrency correctness and value over DFG forecasting remain unverified.
Summary
The paper extends process model forecasting (PMF) — previously limited to directly-follows graphs (DFGs), which cannot represent concurrency — to causal nets. The authors decompose the Fodina discovery algorithm into daily count series over activities, direct-succession relations, length-two loops, and input/output binding patterns, mined from prefix-inclusion sub-logs to avoid look-ahead. Each series is forecast zero-shot with Chronos-2 under a rolling origin (7-day horizon, 1-day stride), and the predicted window totals reconstruct a causal net, converted to a frequency-annotated workflow net and tau-reduced for conformance checking. A partial-trace-aware protocol compares the forecast against re-mined, direct-mined and static references on four public logs (BPI2017, BPI2019, Sepsis, Hospital Billing). Forecasted nets reach conformance close to re-mined references and show clearly higher precision than static discovery on three of four logs (e.g., Align-ETC 0.815 vs 0.487 on BPI2019); on the structurally stable BPI2017 all models score similarly. The binding filter (tauPAT=0.0) improves most metrics but cuts multi-member bindings sharply (36.2% to 1.4% AND share on BPI2019), stripping most modeled concurrency. Main limitations, largely self-acknowledged: an oracle vocabulary built with test-period information, case-complete replay that scores out-of-window behavior, ER that ignores binding frequencies, no ground-truth validation of predicted concurrency, and no DFG-forecasting baseline under the same protocol.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- PMF can be extended from DFGs to concurrency-bearing causal nets by forecasting daily count series of relations and binding patterns and reconstructing under Fodina's construction rules (Sec. 3, 6.1).
- Forecasted causal nets achieve conformance close to re-mined references on four logs (Table 2, e.g., alignment fitness/precision 0.979/0.999 vs 0.988/0.999 on BPI2017 at tauPAT=0.0).
- Static discovery mined once from the training log loses substantial precision on three of four logs (alignment precision 0.487 BPI2019, 0.116 Sepsis, 0.352 Hospital Billing vs Forecast 0.815/0.847/0.977), while remaining competitive on the structurally stable BPI2017 (0.985).
- The binding filter tauPAT=0.0 improves most conformance metrics and makes ER computable, but removes nearly all multi-member (concurrent) bindings, and no evaluated metric can determine whether the remaining concurrency is correct (Sec. 5.3, 6.1).
Strengths
First PMF target beyond DFGs
Extends PMF to causal nets whose input/output bindings encode AND/XOR semantics, addressing the open problem the authors' own prior benchmarking ([26,27]) identified; the decomposition of discovery into forecastable count series (Sec. 3.2) is a reusable design pattern for any count-based miner.
Leakage-aware series extraction
Prefix-inclusion sub-logs (Sec. 3.3, Eq. 3) plus explicit input/output recording rules prevent look-ahead when mining bindings on incomplete daily slices; the output-binding single-record bias is documented rather than hidden.
Unusually careful evaluation protocol
Honest headline about a negative-adjacent finding
Cost and artifact transparency
Weaknesses
No DFG-forecasting baseline under the same protocol
Concurrency is never validated against ground truth
Evaluation uses test-period information
Main analysis rests on optimistic replay
Conditional summaries without inferential statistics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method paper (arXiv 2609.22614, KU Leuven LIRIS) extends process model forecasting (PMF) from directly-follows graphs (DFGs) to causal nets, so that a forecast model can carry explicit concurrency. Four contributions are claimed (Sec. 1): (1) decomposition of Fodina's discovery into daily count series over three relation families plus input/output binding patterns; (2) a prefix-inclusion windowing and forecasting pipeline that reconstructs AND/XOR bindings; (3) a partial-trace evaluation protocol with case-complete replay, tau-reduction, and an entropic-relevance (ER) adaptation for frequency-annotated workflow nets; (4) a four-log study of forecast fidelity, static-baseline drift, and the effect of binding filtering. The abstract's headline results match Table 2: forecast conformance close to re-mined references, clear precision gains over static discovery on three logs, and a filter that improves metrics while destroying concurrency.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is process model forecasting and what does this paper add?
Process model forecasting (PMF) predicts the process model that will govern a future time period, by learning how the components of a mined model evolve over time. Earlier methods forecast directly-follows graphs, which cannot show concurrency. This paper instead forecasts causal nets, whose input/output bindings encode AND (joint) and XOR (alternative) semantics, so the predicted model can contain explicit concurrent behavior.
Does the proposed forecast beat existing approaches?
It is compared against models re-mined from the actual future window (an oracle ablation) and against a static model mined once from the training log. Forecast conformance is close to the re-mined references, and on three of four public logs (BPI2019, Sepsis, Hospital Billing) it is clearly more precise than the static model; on the structurally stable BPI2017 all models perform similarly. No directly-follows forecasting baseline is run under the same protocol.
Is code or data available for this paper?
Yes, as stated in the paper: a GitHub repository (github.com/YongboYu/pmf-concurrency) holds detailed performance results, and Zenodo record 18327515 provides the preprocessed datasets. The method builds on the open Fodina discovery algorithm and the open Chronos-2 time-series foundation model.
What is the main caveat about the reported results?
Evaluation fixes the activity/binding vocabulary and preprocessing using test-period information, replays complete traces that include events outside the forecast window (which the authors call likely optimistic), and the binding filter that improves conformance also removes most modeled concurrency. The paper itself states that no current metric verifies whether the predicted concurrency is correct.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Fig. 1 pipeline diagram, Fig. 2 annotated net) are only assessable via their embedded text labels, not visually.
- Detailed per-window performance tables live in an external GitHub repository, so only the Table 2 aggregates could be reviewed.
- The extracted text has minor typographic artifacts around figure captions, but equations, Table 1 and Table 2 appear complete.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.