Leveraging GPT-4o Efficiency for Detecting Rework Anomaly in Business Processes

Fatemeh Mohammadi; Mohammad Derakhshan; Paolo Ceravolo

REVIEW 4 major objections 7 minor 33 references

Leveraging GPT-4o Efficiency for Detecting Rework Anomaly in Business Processes

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read GPT-4o detects rework anomalies in business event logs with up to 97.94% accuracy, using only one-shot or few-shot plain-text prompts.

desk verdict The GPT-4o rework-detection results are probably real in the sense that the task is easy, but the one/few-shot numbers are compromised by prompt leakage and the missing trivial baseline should worry you. read the letter →

arxiv 2502.06918 v1 pith:3MC22VCC submitted 2025-02-10 cs.LG cs.AI

Mohammad Derakhshan , Paolo Ceravolo , Fatemeh Mohammadi This is my paper · ORCID

classification cs.LGcs.AI

keywords GPT-4oanomalydetectionreworkbusinessprocessminingeventlogslargelanguagemodelspromptengineeringzero-shotone-shotfew-shot

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 claims that GPT-4o-2024-08-06, prompted with plain-text traces of business-process event logs, can detect rework anomalies—tasks repeated unnecessarily within a process instance—at accuracies up to 96.14% for a normal anomaly distribution, 97.94% for a uniform distribution, and 74.21% for an exponential distribution, depending on the prompting strategy. The claim matters because anomaly detection in business processes normally requires trained models, conformance rules, or heavy feature engineering; a frontier LLM with a short prompt would be a low-configuration, accessible alternative. The paper also claims that the distribution of anomalies and the choice of prompting strategy are the main drivers of performance, and that on the same dataset the best GPT-4o results clearly exceed several classical machine-learning baselines.

What carries the argument

The mechanism is the event-variant text transformation plus a prompt template. Each trace is converted into an 'Activity X -> Activity Y -> ...' string, and the human message instructs the model to output 'Variant Id # sequence' for rework anomalies, with zero, one, or three worked examples of repeated activities. The workflow is orchestrated so that the LLM can call a function to fetch the variants, then classify them; the one-shot and few-shot examples are stated to be guaranteed to contain the rework pattern, giving the model a concrete pattern to match within the token limit.

What would settle it

A concrete test is to split the 760 variants so that every trace used in a one-shot or few-shot prompt is held out of the evaluation set, then rerun the three distributions and prompting conditions; if accuracy stays near the reported values, the generalization claim is supported, while a marked drop would show that the model is partly retrieving the example pattern rather than detecting rework.

Watch

Extended reading notes

Core claim

The central discovery is that an LLM, without any fine-tuning, can perform rework-anomaly classification when event traces are formatted as ordered activity sequences such as 'Activity Q -> Activity C -> Activity R -> Activity S' and the model is asked to return the variant IDs containing repeated activity patterns. In the paper's experiments, one-shot prompting with a single example of a repeated activity yields 96.14% accuracy on the normal-distribution dataset; few-shot prompting with three examples yields 97.94% on the uniform-distribution dataset and 74.21% on the exponential-distribution dataset. The authors interpret these results as evidence that LLM pattern recognition, combined with natural-language prompt design, can compete with and exceed traditional machine-learning and conformance approaches for this anomaly type. They also observe a precision/recall trade-off: more examples raise precision but can lower recall, and a strongly skewed anomaly placement degrades performance sharply, with zero-shot prompting timing out entirely.

Load-bearing premise

The load-bearing premise is that the example traces shown in the one-shot and few-shot prompts are not part of the evaluation set; the paper states that the rework pattern is guaranteed to be present in the dataset but never says the example variants are excluded from what is being scored.

Editorial extensions

If this is right

If the claim holds, detecting a rework anomaly requires no training data or feature engineering, sharply lowering the setup cost for process-mining teams.
Prompting technique becomes a tunable control: one-shot and few-shot choices trade precision against recall, so practitioners can select a strategy based on whether false alarms or missed rework is more costly.
Anomaly distribution is as influential as model choice: uniformly scattered anomalies yield near-98% accuracy, whereas bursty placement drops to about 74%, so logs should be checked for placement skew before relying on the method.
On the same synthetic dataset, the best GPT-4o result exceeds all five machine-learning baselines compared in the paper, with a false-discovery rate of 9% under the uniform distribution versus much higher rates for the baselines.

Reading between the lines

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

If the headline accuracies survive a stricter split where the one-shot and few-shot prompt examples are excluded from the evaluation set, then the method likely generalizes beyond the synthetic dataset; if not, the reported numbers partly measure retrieval of the prompt example rather than detection, and the paper never runs that exclusion test.
A natural test the authors do not perform is to vary the number of few-shot examples and the spatial position of injected rework systematically; that would map the performance gap between uniform and exponential distributions to a concrete property of the input ordering.
The same prompt template could probably be adapted to other self-contained anomaly types, such as duplicated resources or missing mandatory activities, since the representation only requires the trace as text; this is an extrapolation beyond the paper's rework-only scope.
Because the dataset is synthetic and loop-free, real logs with loops and long-range dependencies may behave differently; the token limit already forces incremental submission, so splitting long traces could break cross-trace context that the model would otherwise use.

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.

Desk Editor's Note

read the letter

Short version: this is a small empirical paper claiming GPT-4o spots rework anomalies in event logs with 96-98% accuracy under one-shot and few-shot prompting. The headline numbers do not establish generalization, because the one/few-shot prompts contain example variants the authors state are present in the evaluation dataset (Section 5). That is answer leakage. The more interesting numbers are the zero-shot accuracies, which are also high and suggest the task is easy for GPT-4o — or that rework detection is an easy pattern-matching problem.

What is new: this appears to be the first test of GPT-4o-2024-08-06 specifically on rework anomaly detection, across three anomaly distributions and zero/one/few-shot prompting. The writing is clear, and the authors are transparent about the example presence, even if they don't flag it as a threat to validity.

The soft spots, in proportion. One, leakage: the one-shot and few-shot examples are guaranteed to be in the dataset, and there is no disjointness statement. The quantitative impact is bounded — three leaked positive instances move accuracy by less than a point — but it still means those runs are not clean evaluations. Two, the zero-shot exponential row is all zeros because the model timed out, not because it was wrong; treating a timeout as zero accuracy mixes missing data with incorrectness. Three, three runs with no variance; the per-run F1 numbers swing from 19 to 33 on one-shot exponential, so the averaged precision/recall are unstable. Four, the comparison to PCA/IF/DAGMM/TVPM is not apples-to-apples: those baselines come from Shi et al. and were run without the anomaly-distribution protocol. Five, no code or data released.

The missing baseline bothers me most. Rework here is just repeated activities or repeated subsequences. A trivial baseline — flag any variant with a duplicated activity or a repeated length-2 subsequence — would probably nail this synthetic dataset. Without that, the high accuracy doesn't tell us whether GPT-4o's language understanding is doing real work or matching patterns a shell script could catch. The zero-shot results argue for the latter.

Who this is for: people working on LLM applications in BPM/process mining. It deserves a serious referee because the question is live and the flaws are fixable. But I would not accept it as is. The authors need to use held-out prompt examples, report per-run results, handle timeouts as missing, add a simple pattern-matching baseline, and release code/data.

Referee Report

4 major / 7 minor

Summary. The paper proposes a GPT-4o-based tool for detecting rework anomalies in business process event logs, using LangChain and LangGraph to format event variants and prompt the model. It evaluates zero-shot, one-shot, and few-shot prompting on a synthetic dataset of 760 variants under normal, uniform, and exponential anomaly distributions. The reported best accuracies are 96.14% with one-shot prompting on the normal distribution, 97.94% with few-shot prompting on the uniform distribution, and 74.21% with few-shot prompting on the exponential distribution. The paper also compares these results with traditional machine learning methods reported in prior work.

Significance. If the evaluation were sound, this would be a useful demonstration that a general-purpose LLM can detect rework anomalies in event logs with high accuracy and minimal engineering effort, which could lower the barriers to applying anomaly detection in business process management. The paper gives a clear description of a workflow using LangChain and LangGraph, and it builds on a publicly available synthetic dataset. However, the central evidence is compromised by prompt leakage in the one-shot and few-shot conditions and by a problematic handling of timeouts in the zero-shot exponential condition, so the claimed high accuracies cannot be interpreted as estimates of generalization. The comparisons with earlier machine learning baselines are also not made under matched conditions.

major comments (4)

[Section 5, Table 2] The one-shot and few-shot prompts embed example rework patterns that the paper states are 'guaranteed to be present in the dataset.' Because these examples are taken from the evaluation set, the model can detect anomalies by matching the prompt pattern rather than by learning a general rule. The paper does not state that the example variants are disjoint from the test variants, nor does it provide a held-out protocol. Since the headline accuracies are from one-shot and few-shot conditions, the central claim of high detection performance is not established for those settings.
[Section 5 and Table 3] The zero-shot exponential condition reports 0% accuracy for all criteria because the model 'timed out' after five minutes in all three runs. Treating a timeout as zero performance is arbitrary; it conflates API or rate-limit failures with model predictions. The paper should report what happened, e.g., whether the model returned errors or exceeded the time limit, and should exclude or separately analyze such runs. As reported, the zero-shot exponential results are not performance measurements.
[Results, Table 3] Only three runs are performed per condition, and the paper reports averages without standard deviations or confidence intervals. The differences between prompting methods are small in several cases, such as one-shot versus few-shot on the normal distribution (96.14% versus 96.05% accuracy), so the claim that one technique performs best is not statistically supported. The paper should report variance and, where appropriate, significance tests or at least effect sizes.
[Discussion, Table 4] The comparison with LOF, PCA, IF, DAGMM, and TVPM uses results from a prior study that evaluated the same dataset but without accounting for anomaly distributions. The GPT-4o numbers shown are the best results selected across prompt types for each distribution, while the baseline numbers are single reported values from the prior study. This is not an apples-to-apples comparison, and the claim of '46% higher accuracy' over PCA is unsupported by experiments run under matched conditions.

minor comments (7)

[Abstract] The phrase 'onrework anomalies' should be corrected to 'on rework anomalies'.
[Introduction] The statement that the few-shot technique with a uniform distribution yields 'a 46% higher accuracy compared to PCA' is vague; the claim should specify the baseline and the exact accuracy values it refers to.
[Section 5] The sentence 'we repeated each task three times' is ambiguous; the paper should clarify whether each run uses the same 760 variants, the same prompt instances, and whether the three runs are independent draws of the API responses.
[Table 3] The formatting of the table is inconsistent, with some values using commas and others using periods as decimal separators; the table should be aligned and use a single decimal convention.
[Section 7] The paper writes 'GPT 4o' without a hyphen in one instance, which is inconsistent with the spelling 'GPT-4o' used elsewhere.
[References] Reference [5] for LangChain is a preprint; the authors should cite the official LangChain documentation or a more established source.
[Reproducibility] No data or code availability statement is provided; sharing the exact prompts, the variant formatting, and the evaluation script would improve the reproducibility of the experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical prompting benchmark whose central claim is independently supported by zero-shot results; the example-in-dataset disclosure is a test-contamination concern, not a circular derivation.

full rationale

The paper's derivation chain is an empirical evaluation: it feeds event-log variants to GPT-4o under zero-shot, one-shot, and few-shot prompting and reports accuracy. No equation, fitted parameter, or uniqueness theorem is used to derive the reported performance from the prompt content, so none of the enumerated circularity patterns (self-definitional, fitted input called prediction, load-bearing self-citation, imported uniqueness, ansatz via citation, renaming) applies. The self-citations present (e.g., refs. [2], [25], [26], [27], [32]) are background references or the source of the public synthetic dataset; they are not invoked as proof of the model's performance. The one genuine methodological concern is in Section 5: 'In the one-shot and few-shot examples, the rework pattern is guaranteed to be present in the dataset.' This indicates that the hand-written prompt examples correspond to rework patterns that occur in the evaluation data, which risks test-set contamination for the one-shot and few-shot conditions. However, this is an evaluation-validity issue rather than circularity: the prompt examples are not fitted parameters, and the model is not mathematically forced to map those patterns to the correct variant IDs. Moreover, the zero-shot conditions—which contain no example patterns—achieve 94.69% average accuracy under the normal distribution and 96.40% under the uniform distribution, independently supporting the paper's main claim that GPT-4o can detect rework anomalies. The leaked examples are at most three of 760 variants, so even if they are removed, the headline accuracy moves by less than one percentage point. The comparison with Shi et al. baselines may be confounded by different anomaly distributions, but that is a comparability concern, not a circular one. The central result therefore does not reduce to its own inputs by construction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No new theoretical entities are introduced. The evaluation depends on assumptions about dataset representativeness, the definition of rework, and the reliability of output parsing, all of which are unverified.

assumptions (4)

domain assumption The synthetic dataset from Tavares et al. [2] is a valid proxy for real business process event logs.
The entire evaluation is on a synthetic dataset; the paper does not validate on real-world logs, so generalization to practice is assumed.
domain assumption Rework anomalies are defined purely by repetitions of activities (e.g., two X's in a row) and are always undesirable.
The paper's prompt examples define rework as direct repetitions; intentional repetitions with a valid purpose are not considered, so the detection task is simplified.
domain assumption The subset of 760 variants is representative of the original 1,000 variants.
The paper selects 760 variants because of token limits, but does not describe the selection process or show the subset is unbiased.
domain assumption The parsing of GPT-4o's unstructured output into TP/TN/FP/FN is correct.
The paper never specifies how free-text model answers are matched to variant IDs and labels, which is essential for the reported metrics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging GPT-4o Efficiency for Detecting Rework Anomaly in Business Processes." pith.science (2026). https://pith.science/paper/3MC22VCC

@misc{pith2026250206918,
  author       = {Pith},
  title        = {Pith review of: Leveraging GPT-4o Efficiency for Detecting Rework Anomaly in Business Processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3MC22VCC}},
  note         = {Machine review of arXiv:2502.06918}
}

read the original abstract

This paper investigates the effectiveness of GPT-4o-2024-08-06, one of the Large Language Models (LLM) from OpenAI, in detecting business process anomalies, with a focus on rework anomalies. In our study, we developed a GPT-4o-based tool capable of transforming event logs into a structured format and identifying reworked activities within business event logs. The analysis was performed on a synthetic dataset designed to contain rework anomalies but free of loops. To evaluate the anomaly detection capabilities of GPT 4o-2024-08-06, we used three prompting techniques: zero-shot, one-shot, and few-shot. These techniques were tested on different anomaly distributions, namely normal, uniform, and exponential, to identify the most effective approach for each case. The results demonstrate the strong performance of GPT-4o-2024-08-06. On our dataset, the model achieved 96.14% accuracy with one-shot prompting for the normal distribution, 97.94% accuracy with few-shot prompting for the uniform distribution, and 74.21% accuracy with few-shot prompting for the exponential distribution. These results highlight the model's potential as a reliable tool for detecting rework anomalies in event logs and how anomaly distribution and prompting strategy influence the model's performance.

Figures

Figures reproduced from arXiv: 2502.06918 by the authors.

**Figure 1.** Anomaly distribution of datasets. From left to right: normal, uniform, and exponential distribution. The x-axis represents the variant index in the dataset, and the y-axis illustrates the anomaly frequency. For the normal distribution, the dataset is filtered to extract the normal instances, while the abnormal variants are separately identified. We generate a set of insertion indices to integrate these abnormal ins… view at source ↗

**Figure 2.** Anomaly detection workflow The process begins by retrieving the Human message, which is then combined with System messages to initialise the LLM execution. At this stage the LLM assesses whether it needs to call the Setup sub-process. If so, it calls a function from the Setup subprocess to retrieve event variants. The LLM then processes [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Anomaly detection state graph [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 25 canonical work pages

[1]

Hello gpt-4,

OpenAI, “Hello gpt-4,” 2024.https://openai.com/index/hello-gpt-4o/

work page 2024
[2]

M., Oyamada, R

Tavares, G. M., Oyamada, R. S., Barbon, S., & Ceravolo, P. (2023). Trace encoding in process mining: A survey and benchmarking. Engineering Applications of Artifi- cialIntelligence,126,107028. https://doi.org/10.1016/j.engappai.2023.107028

arXiv 2023
[3]

K., Saha, S., Jain, V., Mondal, S., & Chadha, A

Sahoo, P., Singh, A. K., Saha, S., Jain, V., Mondal, S., & Chadha, A. (2024). A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications. https://doi.org/10.48550/arXiv.2402.07927 Title Suppressed Due to Excessive Length 13

arXiv doi:10.48550/arxiv.2402.07927 2024
[4]

N., Park, G., Knopp, B., Graves, N., Rafiei, M., Liß, L., Unterberg, L

Berti, A., Koren, I., Adams, J. N., Park, G., Knopp, B., Graves, N., Rafiei, M., Liß, L., Unterberg, L. T., Zhang, Y., Schwanen, C., Pegoraro, M., & P., W. M. (2024). OCEL (Object-Centric Event Log) 2.0 Specification. ArXiv.https://arxiv.org/ abs/2403.01975

arXiv 2024
[5]

LangChain

Mavroudis, Vasilios. LangChain. 2024.https://www.preprints.org/manuscript/ 202411.0566

arXiv 2024
[6]

Niro, A., & Werner, M. (2024). Detecting Anomalous Events in Object-centric Busi- ness Processes via Graph Neural Networks. https://doi.org/10.48550/arXiv. 2403.00775

work page Pith review arXiv doi:10.48550/arxiv.2403.00775 2024
[7]

& Sungkono, K.R

Sarno, R., Sinaga, F. & Sungkono, K.R. Anomaly detection in business processes using process mining and fuzzy association rule learning. J Big Data 7, 5 (2020). https://doi.org/10.1186/s40537-019-0277-1

doi:10.1186/s40537-019-0277-1 2020
[8]

Rogge-Solti, A., Kasneci, G. (2014). Temporal Anomaly Detection in Business Processes. In: Sadiq, S., Soffer, P., Völzer, H. (eds) Business Process Manage- ment. BPM 2014. Lecture Notes in Computer Science, vol 8659. Springer, Cham. https://doi.org/10.1007/978-3-319-10172-9_15

work page doi:10.1007/978-3-319-10172-9_15 2014

Show all 33 references

[9]

Guan, W., Cao, J., Gao, J., Zhao, H., & Qian, S. (2024). DABL: Detecting Semantic Anomalies in Business Processes Using Large Language Models.https://doi.org/ 10.48550/arXiv.2406.15781

2024 doi
[10]

Minit Process Mining (2o19) Negative Impact of Process Rework on Your Business and the Role of Process Mininghttps://bit.ly/3BfBlD1
[11]

Large language models as general pattern machines

Mirchandani, Suvir, Fei Xia, Pete Florence, Brian Ichter, Danny Driess, Montserrat Gonzalez Arenas, Kanishka Rao, Dorsa Sadigh, and Andy Zeng. Large language models as general pattern machines. arXiv preprint arXiv:2307.04721 2023

2023 arXiv
[12]

An Intelligent LLM-Powered Personalized Assistant for Digital Banking Using LangGraph and Chain of Thoughts

Easin, Arafat Md, Saha Sourav, and Orosz Tamás. An Intelligent LLM-Powered Personalized Assistant for Digital Banking Using LangGraph and Chain of Thoughts. In 2024 IEEE 22nd Jubilee International Symposium on Intelligent Sys- tems and Informatics (SISY), pp. 625-630. IEEE, 20...

2024
[13]

Few-Shot Prompting: Comparison and Examples

Unstructured Data Management Platform, AI Education, Zero-Shot vs. Few-Shot Prompting: Comparison and Examples. https://shelf.io/blog/ zero-shot-and-few-shot-prompting
[14]

Understanding Tokens https://learn.microsoft.com/en-us/ dotnet/ai/conceptual/understanding-tokens

Microsoft Learn. Understanding Tokens https://learn.microsoft.com/en-us/ dotnet/ai/conceptual/understanding-tokens
[15]

(2024, October 6).https://en.wikipedia.org/wiki/GPT-4o

Wikipedia - GPT-4o. (2024, October 6).https://en.wikipedia.org/wiki/GPT-4o

2024
[16]

Singh, V.P. (1998). Normal Distribution. In: Entropy-Based Parameter Estimation in Hydrology. Water Science and Technology Library, vol 30. Springer, Dordrecht. https://doi.org/10.1007/978-94-017-1431-0_5

1998 doi
[17]

Jevremović, V. (2011). Uniform Distribution in Statistics. In: Lovric, M. (eds) International Encyclopedia of Statistical Science. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-04898-2_642

2011 doi
[18]

Singh, V.P. (1998). Exponential Distribution. In: Entropy-Based Parameter Es- timation in Hydrology. Water Science and Technology Library, vol 30. Springer, Dordrecht. https://doi.org/10.1007/978-94-017-1431-0_4

1998 doi
[19]

M., & Fahland, D

Berti, A., Jessen, U., P., W. M., & Fahland, D. (2024). Challenges of Anomaly De- tection in the Object-Centric Setting: Dimensions and the Role of Domain Knowl- edge. ArXiv. https://arxiv.org/abs/2407.09023

2024 arXiv
[20]

Derakhshan, et al

Open AI - Platform usage limit, October 2024https://platform.openai.com/ docs/guides/rate-limits/tier-1-rate-limits 14 M. Derakhshan, et al
[21]

Multi-perspective anomaly detection in business process execution events

Böhmer, Kristof, and Stefanie Rinderle-Ma. Multi-perspective anomaly detection in business process execution events. On the Move to Meaningful Internet Systems: OTM 2016 Conferences: Confederated International Conferences: CoopIS, C&TC, and ODBASE 2016, Rhodes, Greece, October...

2016
[22]

BINet: multivariate busi- ness process anomaly detection using deep learning

Nolle, Timo, Alexander Seeliger, and Max Mühlhäuser. BINet: multivariate busi- ness process anomaly detection using deep learning. International Conference on Business Process Management. Cham: Springer International Publishing, 2018

2018
[23]

Correlation and pattern detection in event networks

Bellandi, V., Ceravolo, P., Maghool, S., Pindaro, M., & Siccardi, S. Correlation and pattern detection in event networks. 2021 IEEE International Conference on Big Data (Big Data). IEEE, 2021

2021
[24]

Bena, Nicola, Marco Anisetti, Gabriele Gianini, and Claudio A. Ardagna. Cer- tifying Accuracy, Privacy, and Robustness of ML-Based Malware Detection. SN Computer Science 5, no. 6, 2024

2024
[25]

J., Ceravolo, P., Damiani, E., Omori, N

Barbon, S. J., Ceravolo, P., Damiani, E., Omori, N. J., & Tavares, G. M. Anomaly Detection on Event Logs with a Scarcity of Labels, 2020 2nd Interna- tional Conference on Process Mining (ICPM), Padua, Italy, 2020, pp. 161-168, doi: 10.1109/ICPM49681.2020.00032

2020
[26]

Are Large Lan- guage Models the New Interface for Data Pipelines?

Barbon Junior, Sylvio, Paolo Ceravolo, Sven Groppe, Mustafa Jarrar, Samira Maghool, Florence Sèdes, Soror Sahri, and Maurice Van Keulen. Are Large Lan- guage Models the New Interface for Data Pipelines?. In Proceedings of the Inter- national Workshop on Big Data in Emergent Di...

2024
[27]

Tuning Machine Learning to Address Process Mining Requirements

Ceravolo, Paolo, Sylvio Barbon, Ernesto Damiani, and Wil Van der Aalst. Tuning Machine Learning to Address Process Mining Requirements. IEEE Access 2024

2024
[28]

Con- formance checking: a state-of-the-art literature review

Dunzer, Sebastian, Matthias Stierle, Martin Matzner, and Stephan Baier. Con- formance checking: a state-of-the-art literature review. In Proceedings of the 11th international conference on subject-oriented business process management, pp. 1-10. 2019
[29]

Conformance checking: foundations, milestones and challenges

Carmona, Josep, Boudewijn van Dongen, and Matthias Weidlich. Conformance checking: foundations, milestones and challenges. In Process mining handbook, pp. 155-190. Cham: Springer International Publishing, 2022

2022
[30]

Global confor- mance checking measures using shallow representation and deep learning

Peeperkorn, Jari, Seppe vanden Broucke, and Jochen De Weerdt. Global confor- mance checking measures using shallow representation and deep learning. Engineer- ing Applications of Artificial Intelligence 123, 2023

2023
[31]

A multi-view deep learning approach for predictive business process mon- itoring

Pasquadibisceglie, Vincenzo, Annalisa Appice, Giovanna Castellano, and Donato Malerba. A multi-view deep learning approach for predictive business process mon- itoring. IEEE Transactions on Services Computing 15, no. 4, 2021

2021
[32]

Tuning Machine Learning to Address Process Mining Requirements

Ceravolo, Paolo, Sylvio Barbon, Ernesto Damiani, and Wil Van der Aalst. Tuning Machine Learning to Address Process Mining Requirements. IEEE Access, 2024

2024
[33]

Shi, Y., Zhang, N., Song, X., Li, H., & Zhu, Q. (2024). Novel approach for industrial process anomaly detection based on process mining. Journal of Process Control, 136, 103165. https://doi.org/10.1016/j.jprocont.2024.103165

2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Hello gpt-4,

OpenAI, “Hello gpt-4,” 2024.https://openai.com/index/hello-gpt-4o/

work page 2024

[2] [2]

M., Oyamada, R

Tavares, G. M., Oyamada, R. S., Barbon, S., & Ceravolo, P. (2023). Trace encoding in process mining: A survey and benchmarking. Engineering Applications of Artifi- cialIntelligence,126,107028. https://doi.org/10.1016/j.engappai.2023.107028

arXiv 2023

[3] [3]

K., Saha, S., Jain, V., Mondal, S., & Chadha, A

Sahoo, P., Singh, A. K., Saha, S., Jain, V., Mondal, S., & Chadha, A. (2024). A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications. https://doi.org/10.48550/arXiv.2402.07927 Title Suppressed Due to Excessive Length 13

arXiv doi:10.48550/arxiv.2402.07927 2024

[4] [4]

N., Park, G., Knopp, B., Graves, N., Rafiei, M., Liß, L., Unterberg, L

Berti, A., Koren, I., Adams, J. N., Park, G., Knopp, B., Graves, N., Rafiei, M., Liß, L., Unterberg, L. T., Zhang, Y., Schwanen, C., Pegoraro, M., & P., W. M. (2024). OCEL (Object-Centric Event Log) 2.0 Specification. ArXiv.https://arxiv.org/ abs/2403.01975

arXiv 2024

[5] [5]

LangChain

Mavroudis, Vasilios. LangChain. 2024.https://www.preprints.org/manuscript/ 202411.0566

arXiv 2024

[6] [6]

Niro, A., & Werner, M. (2024). Detecting Anomalous Events in Object-centric Busi- ness Processes via Graph Neural Networks. https://doi.org/10.48550/arXiv. 2403.00775

work page Pith review arXiv doi:10.48550/arxiv.2403.00775 2024

[7] [7]

& Sungkono, K.R

Sarno, R., Sinaga, F. & Sungkono, K.R. Anomaly detection in business processes using process mining and fuzzy association rule learning. J Big Data 7, 5 (2020). https://doi.org/10.1186/s40537-019-0277-1

doi:10.1186/s40537-019-0277-1 2020

[8] [8]

Rogge-Solti, A., Kasneci, G. (2014). Temporal Anomaly Detection in Business Processes. In: Sadiq, S., Soffer, P., Völzer, H. (eds) Business Process Manage- ment. BPM 2014. Lecture Notes in Computer Science, vol 8659. Springer, Cham. https://doi.org/10.1007/978-3-319-10172-9_15

work page doi:10.1007/978-3-319-10172-9_15 2014