Pith. sign in

REVIEW 4 major objections 4 minor 32 references

SCOPE: Sequential Causal Optimization of Process Interventions

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A backward-induction scheme with causal learners can align sequential process interventions directly from observational logs, without needing a simulator or data augmentation.

desk verdict A clean mapping of backward induction to sequential PresPM with a useful benchmark, but the empirical claim overreaches because the simulators never violate sequential ignorability. read the letter →

arxiv 2512.17629 v4 pith:WDWDGPBH submitted 2025-12-19 cs.LG cs.AI

Jakob De Moor , Hans Weytjens , Johannes De Smedt , Jochen De Weerdt This is my paper
classification cs.LGcs.AI
keywords prescriptiveprocessmonitoringsequentialdecision-makingcausalinferencebackwardinductiondynamictreatmentregimesQ-functionbusinessmanagementKPIoptimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that sequential intervention decisions in business processes can be optimized as a single aligned policy using only historical event logs, without building a simulator or augmenting the data. It claims that combining backward induction with causal effect learners—methods that estimate the KPI under each candidate action from observational data—lets each decision point account for the optimal actions that will be taken later. If correct, organizations could steer running cases toward better outcomes by learning policies that respect how interventions interact over time. The paper is explicit that the method inherits the standard causal assumptions of sequential ignorability, no interference between cases, and positivity, and that violating them would bias the recommendations. It also introduces a reusable semi-synthetic benchmark based on a real-life event log so future sequential PresPM methods can be compared.

What carries the argument

The central object is the pair of Q-function and value function defined recursively from the last decision point K back to the first: Q_k(prefix, action) is the expected future KPI if action is taken now and optimal actions are taken afterward; V_k(prefix) is the maximum over actions of Q_k. Under sequential ignorability, SUTVA, and positivity, these are identifiable from observational logs, and the optimal action at each point is the argmax of Q_k. The paper estimates Q_k with causal learners (S-, T-, or RA-learner) and uses the regret-based value function to build training targets for the preceding decision point, letting estimation error propagate less through the chain.

What would settle it

Run the method on an observational log where an unmeasured confounder is known to exist—e.g., a process where case priority is hidden from the log but drives both escalation decisions and outcome—and compare the recommended policy against a randomized trial or a policy learned with the hidden variable included. If the recommendations deviate from the true optimal policy in the direction of the confounder bias, the claim that SCOPE identifies the optimal sequential policy from observational data is falsified. A simpler check: on a synthetic process that violates sequential ignorability, SCOPE's

Watch

Extended reading notes

Core claim

The paper's central claim is that the optimal sequential intervention policy is identifiable from observational data through backward induction, and that this policy can be estimated in practice by using a causal learner to fit the Q-function at each decision point. Starting at the final decision point, the method estimates the expected KPI for each action given the case prefix; the value function then propagates the best achievable outcome one step back, so that an earlier action is evaluated by the KPI it leads to under later optimal choices. A regret-based form of the value function (V_k = E[V_{k+1} + Q(optimal) - Q(observed)]) makes the procedure more robust to model misspecification. Th

Load-bearing premise

The load-bearing premise is sequential ignorability: that every prefix contains all confounders that influence both the intervention choice and the final KPI, so that the causal learners can estimate unbiased Q-functions from observational data.

Editorial extensions

If this is right

  • Sequential intervention policies can be learned directly from historical logs, avoiding the reality gap of MDP approximations or data augmentation.
  • The advantage over independent per-decision-point optimization grows with the number of decision points, because aligned policies exploit inter-step dependencies.
  • Performance improves with larger training sets, since error accumulation in the backward chain decreases.
  • The released semi-synthetic benchmark (SimBPIC17) gives future work a standardized setting for sequential PresPM evaluation.
  • Under strong confounding, the RA-learner variant degrades least, suggesting meta-learner choice matters for real-world logs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If sequential ignorability fails in a real process—e.g., unmeasured case-level variables drive both intervention choice and outcome—the learned policy will be biased no matter how good the base models are; the paper acknowledges this and suggests instrumental variables as a future direction but does not test it.
  • Because the regret-based formulation only estimates the difference between optimal and observed actions' Q-values, in settings where the historical policy is already near-optimal the signal for improvement may be weak, so gains reported on synthetic logs with suboptimal bank policies may overstate real-world uplift.
  • The method models dependencies within a case but assumes no interference between cases; extending it to shared-resource settings would require a different identification strategy.
  • A direct extension would be to handle continuous or multi-valued action spaces beyond the small categorical sets tested, though this would require new causal learners.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes SCOPE, a method for prescriptive process monitoring (PresPM) that learns sequential intervention policies via backward induction combined with causal meta-learners (S-, T-, RA-learners). It uses a regret-based value function to propagate estimated Q-values from the final decision point back to earlier ones, avoiding the need for process simulation or data augmentation. Experiments on a synthetic bank simulator (SimBank) and a new semi-synthetic simulator based on the BPIC17 event log (SimBPIC17) compare SCOPE with independent per-decision-point optimization (SEP) and an MDP/Q-learning baseline (KMeans-Q). The results indicate that SCOPE generally outperforms these baselines, with the gap growing as the number of decision points increases. The authors release their code and the SimBPIC17 benchmark as a reusable resource.

Significance. If the result holds, SCOPE provides a practical, simulation-free alternative to RL-based sequential PresPM that works directly on observational event logs. The combination of backward induction with causal meta-learners is well-motivated, and the regret-based value function is a sensible design choice with some theoretical grounding. The release of the new SimBPIC17 benchmark and the code is a useful contribution to the community. However, the empirical claim is currently broader than the evidence supports: both simulators satisfy sequential ignorability by construction, the KMeans-Q baseline is re-implemented with modified tuning, and the paper itself reports settings where SCOPE does not outperform the baselines. These issues limit the external validity of the headline claim.

major comments (4)
  1. [Abstract and Section 4.5, Figs. 1, 2, 4] The abstract and Section 4.5 state that SCOPE 'consistently outperforms state-of-the-art PresPM techniques', but the paper's own results contain exceptions: SimBank (1K, δ=0.99) in Fig. 1, SimBPIC17 (δ=0.99, 2 decision points) in Fig. 4, and the T-learner (δ=0.99) case in Fig. 2. These exceptions should be explicitly acknowledged and the wording tempered to 'generally outperforms', or the claim qualified in terms of the specific settings.
  2. [Section 3.2, 4.1, 4.6] The identification of the optimal policy relies on sequential ignorability. In both simulators, the historical action policy is generated from prefix features that are included in the models, so the training data satisfy ignorability by construction. Varying δ only changes the fraction of random actions, not the presence of unmeasured confounders. The limitation paragraph in Section 4.6 concedes this possibility, but the final sentence claims that experiments show SCOPE works under 'realistic levels of confounding'—a claim not supported by the experiments. The authors should either add an experiment with an unmeasured confounder (e.g., a latent variable affecting both action and outcome) or restrict the empirical claim to the ignorable setting.
  3. [Section 4.2, footnote 9] The KMeans-Q baseline is not evaluated with the original tuning protocol from Branchi et al. Because the authors found the original sequential tuning to perform poorly, they jointly tuned KMeans and the RL component with a combined silhouette+reward metric and gave it twice the tuning budget. This is a reasonable modification, but it means the comparison is not a direct test of the published method. Please report the original tuning result or at least ablate the impact of this modification so readers can gauge how much of the gap is due to the algorithmic change rather than tuning.
  4. [Section 3.2, Eq. (5)] The pseudo-outcome formula for the RA-learner appears problematic. For an action ak different from the observed action fk, the formula adds (Q̂(σ,fk)-y) and (Q̂(σ,fk)-Q̂(σ,bk)); as written, this is not a standard meta-learner pseudo-outcome and seems to double-count the baseline action. If this is a known formulation from Acharki et al. [2], please provide the exact reference and correct any transcription errors; otherwise the formula should be fixed. This matters because RA-learner results are reported in Fig. 2.
minor comments (4)
  1. [Figure 1 caption] The caption has a typo: 'SCOPEand Sep' should be 'SCOPE and SEP'.
  2. [Eq. (5)] The notation Φ(ak)k should likely be Φ_k(ak) for readability; also define the baseline action more clearly.
  3. [Section 4.4] The random policy baseline is reported as -132.597 ± 0.983 in Fig. 1, but its definition (uniform over action sequences? independent per decision point?) and whether it is identical across all settings are not clearly explained.
  4. [Section 4.1] The description of SimBPIC17 states that 'the throughput time ... is sampled from a uniform distribution' and that cost constants and the bank policy threshold are fixed. Please provide the actual values (or a table) so that the benchmark is reproducible without inspecting the code.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; the central equations are self-contained, though the empirical evaluation leans on self-authored benchmarks.

full rationale

I walked the derivation chain and found no step where a claimed prediction is equivalent to an input by construction. The backward-induction recurrences (Eqs. 1-4 and Algorithm 1) are standard dynamic-programming recursions: Q_k is learned by regressing the next value target on the prefix and action, and V_k is computed from Q_opt - Q_obs. No fitted parameter is renamed as a prediction; the causal learners are trained on generated logs and scored by gain over the historical policy. Identifiability is attributed to external DTR literature [21], not to an author-unique theorem. The main concern is external validity rather than circularity: SimBank is the authors' own simulator [10], and SimBPIC17 is introduced in this paper with hand-coded inter-decision interactions (call effects depending on average duration) that are well aligned with what SCOPE is designed to exploit. Section 4.6 also concedes that sequential ignorability may not hold in real business processes. These are benchmark-independence and generalization limitations, not a logical reduction of the method's results to its assumptions. I therefore report a minor score for the self-cited simulation evidence, but no substantive circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method relies on standard causal identifiability conditions and on the representativeness of two author-built simulators. SimBPIC17's constants and bank threshold are hand-chosen and appear in the paper without values, making the benchmark underspecified as published.

free parameters (3)
  • SimBPIC17 cost constants cost_tpt and cost_call = not reported
    Define the KPI trade-off in the new semi-synthetic benchmark (Section 4.1); hand-chosen by the authors and not justified from data.
  • SimBPIC17 bank policy threshold and effect modifiers = 4025 time units; loan types car/loan takeover
    Set the confounding mechanism and effect heterogeneity in the simulator (Section 4.1); these choices shape how much gain SCOPE can achieve over the bank policy.
  • SimBPIC17 throughput-time sampling distribution = uniform, parameters unspecified
    The KPI uses tpt_files sampled from a uniform distribution (Section 4.1); bounds are not given, so the benchmark is underspecified in the paper.
assumptions (5)
  • domain assumption Sequential ignorability: no unmeasured confounders given the prefix at each decision point
    Stated in Section 3.2 and in the limitations; without it the Q-functions estimated by causal learners are biased.
  • domain assumption SUTVA: no interference across cases and well-defined treatment versions
    Stated in Section 3.2 and limitations; makes potential outcomes under a policy well-defined per case.
  • domain assumption Positivity: every candidate action has positive probability for every relevant prefix
    Needed to estimate Q_k for all actions from observational data; noted in Section 3.2.
  • ad hoc to paper SimBank and SimBPIC17 are representative of real business process dynamics
    The empirical claim is evaluated only on these author-built simulators; if their outcome functions do not generalize, the claimed consistent outperformance may not hold in practice.
  • standard math Backward induction identifies the optimal policy when Q-functions are known
    Used throughout Section 3.2; this is standard dynamic programming.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCOPE: Sequential Causal Optimization of Process Interventions." pith.science (2026). https://pith.science/paper/WDWDGPBH

@misc{pith2026251217629,
  author       = {Pith},
  title        = {Pith review of: SCOPE: Sequential Causal Optimization of Process Interventions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDWDGPBH}},
  note         = {Machine review of arXiv:2512.17629}
}
read the original abstract

Prescriptive Process Monitoring (PresPM) recommends interventions during running business processes to optimize key performance indicators (KPIs). In realistic settings, interventions are rarely isolated: organizations need to align sequences of interventions to jointly steer the outcome of a case. Existing PresPM approaches only partially address this challenge. Many focus on a single intervention decision, while others treat multiple interventions independently, ignoring how they interact over time. Methods that do address these dependencies depend either on simulation or data augmentation to approximate the process to train a Reinforcement Learning (RL) agent, which may create a reality gap and introduce bias. We introduce SCOPE (Sequential Causal Optimization of Process Interventions), a PresPM approach that learns aligned sequential intervention recommendations. SCOPE employs backward induction to estimate the effect of each candidate intervention action, propagating its impact from the final decision point back to the first. By leveraging causal learners, our method can utilize observational data directly, unlike methods that require constructing process approximations for RL. Experiments on both an existing synthetic dataset and a new semi-synthetic dataset show that SCOPE consistently outperforms state-of-the-art PresPM techniques in optimizing the KPI. The novel semi-synthetic setup, based on a real-life event log, is provided as a reusable benchmark for future work on sequential PresPM.

Figures

Figures reproduced from arXiv: 2512.17629 by the authors.

Figure 1
Figure 1. Gain on SimBank across confounding levels (δ) for different training sizes. The shaded area shows one standard error over 10 iterations. SCOPE and Sep: S-learner with XGBoost. 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 140 160 180 200 220 240 Gain over bank policy (%) S-Learner 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 T-Learner 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 RA-Learner SCOPE (ours)… view at source ↗
Figure 2
Figure 2. Gain on SimBank across confounding levels (δ) for different learners. The shaded area shows one standard error over 10 iterations. SCOPE and Sep: XGBoost, trained on 10K cases. 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 50 100 150 200 250 Gain over bank policy (%) XGBoost 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 Random Forest 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 MLP/LSTM SCOPE (ours) SEP… view at source ↗
Figure 3
Figure 3. Gain on SimBank across confounding levels (δ) for different base models. The shaded area shows one standard error over 10 iterations. SCOPE and Sep: S-learner trained on 10K cases [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Gain on SimBPIC17 across varying numbers of decision points for three confounding levels (δ). The shaded area shows one standard error over 10 iterations. SCOPE and Sep: S-learner with XGBoost, trained on 10K cases. 4.6 Discussion In Summary, the setup of SCOPE demonst…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 10 canonical work pages

  1. [2]

    In: Proceedings of the 40th International Conference on Machine Learning

    Acharki, N., Lugo, R., Bertoncello, A., Garnier, J.: Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects. In: Proceedings of the 40th International Conference on Machine Learning. ICML’23, JMLR.org (2023)

    2023

  2. [1]

    Abbasi, M., Khadivi, M., Ahang, M., Lasserre, P., Lucet, Y., Najjaran, H.: An inno- vative data-driven and adaptive reinforcement learning approach for context-aware prescriptive process monitoring (2025),https://arxiv.org/abs/2501.10543

    arXiv 2025

  3. [3]

    Bozorgi, Z.D., Dumas, M., Rosa, M.L., Polyvyanyy, A., Shoush, M., Teinemaa, I.: Learning when to treat business processes: Prescriptive pro- cess monitoring with causal inference and reinforcement learning (2023), https://arxiv.org/abs/2303.03572

    arXiv 2023

  4. [4]

    In: 2021 3rd International Conference on Process Mining (ICPM)

    Bozorgi, Z.D., Teinemaa, I., Dumas, M., Rosa, M.L., Polyvyanyy, A.: Pre- scriptive process monitoring for cost-aware cycle time reduction. In: 2021 3rd International Conference on Process Mining (ICPM). pp. 96–103 (2021). https://doi.org/10.1109/ICPM53251.2021.9576853

    arXiv 2021

  5. [5]

    Branchi, S., Buliga, A., Francescomarino, C.D., Ghidini, C., Meneghello, F., Ronzani, M.: Recommending the optimal policy by learning to act from temporal data (2023),https://arxiv.org/abs/2303.09209

    arXiv 2023

  6. [6]

    In: Di Ciccio, C., Dijkman, R., del Río Ortega, A., Rinderle-Ma, S

    Branchi, S., Di Francescomarino, C., Ghidini, C., Massimo, D., Ricci, F., Ronzani, M.: Learning to act: A reinforcement learning approach to recommend the best next activities. In: Di Ciccio, C., Dijkman, R., del Río Ortega, A., Rinderle-Ma, S. (eds.) Business Process Management Forum. pp. 137–154. Springer International Publishing, Cham (2022)

    2022

  7. [7]

    European Journal of Operational Research328(2), 620–632 (2026).https://doi.org/https: //doi.org/10.1016/j.ejor.2025.09.015,https://www.sciencedirect

    Caljon, D., Van Belle, J., Berrevoets, J., Verbeke, W.: Optimizing treat- ment allocation in the presence of interference. European Journal of Operational Research328(2), 620–632 (2026).https://doi.org/https: //doi.org/10.1016/j.ejor.2025.09.015,https://www.sciencedirect. com/science/article/pii/S0377221725007325

    work page doi:10.1016/j.ejor.2025.09.015 2026

  8. [8]

    In: Banerjee, A., Fukumizu, K

    Curth, A., van der Schaar, M.: Nonparametric estimation of heteroge- neous treatment effects: From theory to learning algorithms. In: Banerjee, A., Fukumizu, K. (eds.) Proceedings of The 24th International Confer- ence on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 130, pp. 1810–1818. PMLR (13–15 Apr 2021), https...

    2021

Show all 32 references
  1. [9]

    Information Systems116(2023).https://doi.org/10.1016/j.is.2023.102198

    Dasht Bozorgi, Z., Teinemaa, I., Dumas, M., La Rosa, M., Polyvyanyy, A.: Prescriptive process monitoring based on causal effect estimation. Information Systems116(2023).https://doi.org/10.1016/j.is.2023.102198

    2023

  2. [10]

    In: Senderovich, A., Cabanillas, C., Vanderfeesten, I., A

    De Moor, J., Weytjens, H., De Smedt, J., De Weerdt, J.: Simbank: From simulation to solution in prescriptive process monitoring. In: Senderovich, A., Cabanillas, C., Vanderfeesten, I., A. Reijers, H. (eds.) Business Process Management Forum. pp. 165–182. Springer Nature Switze...

    2026

  3. [11]

    van Dongen, B.: Bpi challenge (2017).https://doi.org/10.4121/uuid: 5f3067df-f10b-45da-b98b-86ae4c7a310b

    2017 doi

  4. [12]

    Journal of the American Statistical Association81(396), 945–960 (1986),http://www.jstor.org/stable/2289064 16 De Moor et al

    Holland, P.W.: Statistics and causal inference. Journal of the American Statistical Association81(396), 945–960 (1986),http://www.jstor.org/stable/2289064 16 De Moor et al

    1986

  5. [13]

    Statistics in Medicine34(26), 3424–3443 (Nov 2015).https://doi.org/10.1002/sim.6558, epub 2015 Jun 21

    Huang, X., Choi, S., Wang, L., Thall, P.F.: Optimization of multi-stage dynamic treatment regimes utilizing accumulated data. Statistics in Medicine34(26), 3424–3443 (Nov 2015).https://doi.org/10.1002/sim.6558, epub 2015 Jun 21

    2015 doi

  6. [14]

    Journal of the Royal Statistical Society Series C: Applied Statistics72(2), 434–449 (03 2023).https://doi.org/ 10.1093/jrsssc/qlad016,https://doi.org/10.1093/jrsssc/qlad016

    Illenberger, N., Spieker, A.J., Mitra, N.: Identifying optimally cost-effective dynamic treatment regimes with a q-learning approach. Journal of the Royal Statistical Society Series C: Applied Statistics72(2), 434–449 (03 2023).https://doi.org/ 10.1093/jrsssc/qlad016,https://d...

    2023 doi

  7. [15]

    In: Proceedings of the 31st ACM International Conference on Information & Knowledge Management

    Jiang, S., Sun, Y.: Estimating causal effects on networked observational data via representation learning. In: Proceedings of the 31st ACM International Conference on Information & Knowledge Management. p. 852–861. CIKM ’22, Association for Computing Machinery, New York, NY, U...

    2022

  8. [16]

    https://doi.org/10.7717/peerj-cs.1097

    Kubrak, K., Milani, F., Nolte, A., Dumas, M.: Prescriptive process monitoring: Quo vadis? PeerJ Computer Science8, e1097 (2022). https://doi.org/10.7717/peerj-cs.1097

    2022 doi

  9. [17]

    Proceedings of the National Academy of Sciences116(2017).https://doi.org/10.1073/pnas.1804597116

    Künzel, S., Sekhon, J., Bickel, P., Yu, B.: Meta-learners for estimating hetero- geneous treatment effects using machine learning. Proceedings of the National Academy of Sciences116(2017).https://doi.org/10.1073/pnas.1804597116

    2017 doi

  10. [18]

    In: 2020 2nd International Conference on Process Mining (2020)

    Leoni, M.d., Dees, M., Reulink, L.: Design and evaluation of a process-aware recommender system based on prescriptive analyt- ics. In: 2020 2nd International Conference on Process Mining (2020). https://doi.org/10.1109/ICPM49681.2020.00013

    2020

  11. [19]

    Emerging Themes in Epidemiology15(2018)

    Lousdal, M.L.: An introduction to instrumental variable assumptions, val- idation and estimation. Emerging Themes in Epidemiology15(2018). https://doi.org/10.1186/s12982-018-0069-7

    2018 doi

  12. [20]

    Journal of Educational Psychology66(5), 688–701 (1974)

    Rubin, D.B.: Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology66(5), 688–701 (1974). https://doi.org/10.1037/h0037350,https://doi.org/10.1037/h0037350

    1974 doi

  13. [21]

    Statistical Science29(4), 640–661 (Nov 2014).https://doi.org/10.1214/13-STS450

    Schulte, P.J., Tsiatis, A.A., Laber, E.B., Davidian, M.: Q- and a-learning methods for estimating optimal dynamic treatment regimes. Statistical Science29(4), 640–661 (Nov 2014).https://doi.org/10.1214/13-STS450

    2014 doi

  14. [22]

    Shalit, U., Johansson, F.D., Sontag, D.: Estimating individual treatment effect: gen- eralization bounds and algorithms (2017),https://arxiv.org/abs/1606.03976

    2017 arXiv

  15. [23]

    Künstliche Intelligenz 39, 119–140 (2025).https://doi.org/10.1007/s13218-024-00881-6, https://doi.org/10.1007/s13218-024-00881-6

    Shoush, M., Dumas, M.: Prescriptive process monitoring under resource constraints: A reinforcement learning approach. Künstliche Intelligenz 39, 119–140 (2025).https://doi.org/10.1007/s13218-024-00881-6, https://doi.org/10.1007/s13218-024-00881-6

    2025 doi

  16. [24]

    Shoush, M., Dumas, M.: Intervening with confidence: Conformal prescriptive monitoring of business processes (2022),https://arxiv.org/abs/2212.03710

    2022 arXiv

  17. [25]

    In: Process Mining Workshops

    Shoush, M., Dumas, M.: Prescriptive process monitoring under resource constraints: A causal inference approach. In: Process Mining Workshops. Springer (2022)

    2022

  18. [26]

    Data & Knowledge Engineering (2025)

    Shoush, M., Dumas, M.: White box specification of intervention policies for prescriptive process monitoring. Data & Knowledge Engineering (2025)

    2025

  19. [27]

    In: Business Process Management Forum

    Teinemaa,I.,Tax,N.,deLeoni,M.,Dumas,M.,Maggi,F.M.:Alarm-basedprescrip- tive process monitoring. In: Business Process Management Forum. Springer (2018)

    2018

  20. [28]

    ACM Trans

    Verenich, I., Dumas, M., Rosa, M.L., Maggi, F.M., Teinemaa, I.: Survey and cross-benchmark comparison of remaining time prediction methods in business process monitoring. ACM Trans. Intell. Syst. Technol.10(4) (Jul 2019). https://doi.org/10.1145/3331449,https://doi.org/10.1145...

    2019 doi

  21. [29]

    In: Business Process Management Forum

    Weinzierl, S., Dunzer, S., Zilker, S., Matzner, M.: Prescriptive business process monitoring for recommending next best actions. In: Business Process Management Forum. Springer (2020)

    2020

  22. [30]

    reinforcement learning

    Weytjens, H., Verbeke, W., De Weerdt, J.: Timed process interventions: Causal inference vs. reinforcement learning. In: Business Process Management Workshops. Springer (2024)

    2024

  23. [31]

    Annals of Applied Statistics 19(2), 1426–1447 (June 2025).https://doi.org/10.1214/24-AOAS1984, https://doi.org/10.1214/24-AOAS1984

    Zhan, Z., Liu, Z., Lin, C., Yi, D., Liu, J., Yang, Y.: Censored c-learning for dynamic treatment regime in colorectal cancer study. Annals of Applied Statistics 19(2), 1426–1447 (June 2025).https://doi.org/10.1214/24-AOAS1984, https://doi.org/10.1214/24-AOAS1984

    2025 doi

  24. [32]

    Biometrics74(3), 891–899 (September 2018).https://doi.org/10.1111/biom.12836, https://doi.org/10.1111/biom.12836, epub 2017 Dec 11, PMID: 29228509

    Zhang, B., Zhang, M.: C-learning: A new classification framework to estimate optimal dynamic treatment regimes. Biometrics74(3), 891–899 (September 2018).https://doi.org/10.1111/biom.12836, https://doi.org/10.1111/biom.12836, epub 2017 Dec 11, PMID: 29228509

    2018 doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.