Pith. sign in

REVIEW 3 major objections 3 minor 15 references

FairLoop: Software Support for Human-Centric Fairness in Predictive Business Process Monitoring

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper presents FairLoop, a tool that distills a neural prediction model into an editable decision tree and lets users delete unfairly biased rules before re-labeling data and fine-tuning the model for fairer predictions.

desk verdict FairLoop is a useful tool demo with an honest limitations section, but its central fairness claim is unmeasured—worth a referee who will push for either softened wording or one quantitative check. read the letter →

arxiv 2508.20021 v1 pith:JOP5GV75 submitted 2025-08-27 cs.LG

Felix Möhrlein , Martin Käppel , Julian Neuberger , Sven Weinzierl , Lars Ackermann , Martin Matzner , Stefan Jablonski This is my paper
classification cs.LG
keywords predictivebusinessprocessmonitoringfairnessdecisiontreedistillationhuman-in-the-loopbiasmitigationneuralnetworksexplainableAImining
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

FairLoop is a graphical tool that gives users a way to fix unfair predictions in a process-monitoring neural network without retraining from scratch or banning a sensitive attribute outright. It turns the network's decision logic into an editable decision tree, lets a user delete rules that encode negative bias, then uses the corrected tree to re-label the training data and fine-tune the network. The work's aim is to make context-aware fairness practical: the same attribute (e.g., gender) can stay in the model where it is clinically or operationally necessary while being removed where it is discriminatory. The paper demonstrates the workflow on a simulated cancer-screening event log and argues the approach fills a gap left by uniform bias-removal methods. It does not claim quantitative fairness results; it claims the tool makes selective, human-guided fairness intervention possible.

What carries the argument

The editable decision tree: an interpretable tree trained on the set of all event-log prefixes and the MLP's predictions for them, converted into a mutable structure that users can alter by deleting nodes or retraining subtrees. It is the translation layer between the black-box MLP and the human; after the user edits it, its predictions serve as the new target labels for fine-tuning, which is what propagates the edit back into the neural model. The repeated distill-alter-tune cycle is the loop that drives FairLoop.

What would settle it

Measure the agreement rate between the distilled decision tree and the MLP on all event-log prefixes; then delete one biased rule, fine-tune, and check whether the MLP's predictions on prefixes matching that rule actually shift relative to a randomly chosen control set. If the agreement is low or the targeted predictions do not move, the loop has failed to transfer the edit.

Watch

Extended reading notes

Core claim

The paper's central claim is that a human-in-the-loop cycle—distilling a decision tree from a trained neural network, letting a user delete rules that encode negative bias, re-labeling every event prefix with the edited tree's predictions, and fine-tuning the neural network on that re-labeled data—can remove unfair behavior selectively while preserving the model's predictive expressiveness and its legitimate use of the same sensitive attribute in other contexts. The paper demonstrates this in a browser-based tool on a simulated cancer-screening event log, showing where gender drives a fair screening decision (positive bias) and where it drives an unfair refusal (negative bias). It argues tha

Load-bearing premise

The distilled decision tree must faithfully reproduce the MLP's decisions on the relevant prefixes; if it does not, deleting a rule from the tree will not change the MLP's behavior in the intended way.

Editorial extensions

If this is right

  • A user can remove one discriminatory rule, such as 'gender = female → refuse screening,' while keeping other rules that use gender for clinically necessary decisions.
  • Because the final predictor is still the MLP, the interpretability of the tree does not come at the cost of the model's predictive capacity.
  • The iterative loop lets users re-distill after fine-tuning to see whether the edited behavior took effect and to make further adjustments.
  • The tool lowers the technical barrier for fairness intervention: a domain expert with no programming experience can do the editing in a browser.

Reading between the lines

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

  • The same distill-edit-retrain recipe should transfer to other black-box models and other prediction targets (e.g., outcome or remaining-time prediction) whenever the distilled tree is a faithful enough mirror of the model; the paper leaves that fidelity unmeasured.
  • Without a built-in fairness metric, users currently judge success by inspecting the tree; adding an automatic before/after comparison of prediction rates per sensitive group would turn the loop from a manual check into a measurable intervention.
  • If real event logs with sensitive attributes produce larger, messier trees than this simulated log, the human-editing step may become the bottleneck; guiding user attention to the most biased subtrees would then be necessary.
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

3 major / 3 minor

Summary. The paper presents FairLoop, a web-based tool for human-in-the-loop bias mitigation in predictive business process monitoring (PBPM). The proposed workflow has four steps: train an MLP on an event log for next-activity prediction; distill a decision tree from the MLP using all prefixes in the event log and their MLP predictions; let a domain expert inspect the tree and remove or modify rules judged to encode negative bias (while preserving positive, context-dependent uses of sensitive attributes); and re-label the training data with the edited tree, then fine-tune the original MLP on the relabeled data. The authors argue that this selective, context-aware intervention improves on fairness methods that remove sensitive attributes uniformly. The paper includes an architecture description, a simulated cancer-screening demonstration event log, a screenshot of the interactive tree editor, and links to public code, a video, and a live demo. The authors explicitly state that no user study has been conducted and that the tool currently includes no built-in fairness metrics; the demonstration is assessed by manual inspection of the distilled tree.

Significance. If the central mechanism works, FairLoop would be a practically useful contribution: it offers a graphical, human-guided workflow that lets PBPM users inspect black-box MLP decisions through distilled trees and apply context-sensitive fairness interventions rather than blanket removal of sensitive attributes. The modular architecture and public code are strengths, and the paper is honest about its current limitations. However, the load-bearing claim—that the distilled tree faithfully represents the MLP and that editing it and fine-tuning on relabeled data actually removes bias—is not quantitatively demonstrated anywhere in the manuscript. The paper reports no distillation fidelity, no fairness metric, no before/after prediction comparison, and no accuracy/F1 figures, despite stating that such metrics are shown in the interface. The significance of the contribution therefore remains conditional: the tool is plausible and demonstrable, but the evidence supplied does not substantiate the claim that FairLoop enables fairer predictions.

major comments (3)
  1. [Section 2, Distillation step] The decision tree is built from "all possible prefixes in the event log and the corresponding prediction made by the MLP," but no measure of distillation fidelity is reported. The entire intervention depends on the tree being a faithful surrogate of the MLP: if the tree misrepresents the MLP's decision logic, then removing a rule such as "gender = female -> refuse screening" may alter a branch the MLP does not actually use, and the relabeled training data will not reflect the intended change. Please report tree–MLP agreement on held-out prefixes (and, if available, tree size/depth or fidelity by output class), and show the before/after distilled trees for the simulated log to demonstrate that the edit changes the surrogate in the intended place.
  2. [Section 4, Limitations and demonstration] The paper's abstract and Section 1 claim that the re-labeling and fine-tuning procedure is used "to eliminate the bias," but Section 4 explicitly notes the "absence of built-in metrics for quantifying the fairness of predictive models" and that interventions are assessed "solely through manual inspection of the distilled decision tree." The only demonstration is a screenshot (Fig. 3); no quantitative before/after comparison is provided. Please add a concrete fairness evaluation for the simulated cancer-screening log—e.g., demographic parity or equalized odds of the next-activity prediction with respect to gender, before and after the intervention—together with the accuracy/F1/precision/recall numbers that Section 2 says are displayed. Without such measurements, the central claim that FairLoop removes bias is unsupported.
  3. [Section 2, Re-labeling and fine-tuning] Even if the tree is an accurate surrogate, the pipeline has a potential circularity: the tree is derived from the MLP, the edited tree generates new labels, and the MLP is fine-tuned on those labels. The MLP may converge to the edited behavior largely by construction on the training prefixes, but it is not shown whether the fine-tuned model generalizes the edit to unseen prefixes or merely memorizes the relabeled training set. It is also not shown whether the model simply redirects its decision to correlated features (e.g., other attributes correlated with gender). Please report fine-tuning details (epochs, learning rate, data split) and a holdout evaluation showing that predictions on affected prefixes actually change, plus an ablation comparing fine-tuning on edited-tree labels versus simply retraining without the sensitive attribute.
minor comments (3)
  1. [Section 1, paragraph 4] Typo: "from even log data" should be "from event log data."
  2. [Figure 1] The numbered steps 1–8 in the architecture diagram are not explained in the text. A short caption or corresponding paragraph enumerating the steps would make the workflow much easier to follow.
  3. [Section 2, decision tree construction] The phrase "all possible prefixes" is ambiguous for a process log: does this include prefixes truncated at every event position, with all case attributes and encoding variants? Clarify the exact construction and how the MLP output is mapped to a discrete target for tree training. Also, the "custom implementation" of the decision tree is not described; please specify what operations (removing a node, retraining a subtree) are supported and how they are implemented.

Circularity Check

2 steps flagged · score 6.0 of 10

Bias-removal outcome is generated by the same relabeling step used to evaluate it; no independent fairness metric breaks the loop.

  1. self definitional [Section 2, System Overview (re-labeling and fine-tuning paragraph)]
    "To this end, we assign each prefix in the event log the target predicted by the (now fairer) decision tree, which in turn lets the MLP learn to make fairer predictions during fine-tuning."

    The 'fairer' decision tree is exactly the tree the user has just edited; 'fairer predictions' are therefore defined as whatever that edited tree outputs. The MLP is then fine-tuned to match those self-generated labels, so any later display of 'fairer' behavior is not an independent empirical outcome but the direct result of training on labels derived from the same artifact. The paper confirms that the only assessment is 'manual inspection of the distilled decision tree' (Section 4), which is a distillation of a model trained on those labels. Thus the claimed bias removal is built into the relabeling step rather than demonstrated on external data or metrics.

  2. fitted input called prediction [Section 2, System Overview; Section 4, Limitations]
    "After this step, the user can view the effects of their alterations by distilling a new decision tree from the fine-tuned MLP model. ... Another key limitation is the absence of built-in metrics for quantifying the fairness of predictive models, meaning the effectiveness of user-driven interventions is currently assessed solely through manual inspection of the distilled decision tree."

    The 'effect' of the alterations is observed by distilling a tree from a model that was fine-tuned on labels supplied by the altered tree. Since the training objective was to fit those labels, a high-fidelity distillation will simply reproduce the edited decision logic; it is not a hold-out validation of bias removal. The paper's own limitation statement confirms there is no external fairness metric, so the assessment loop is closed and the apparent success is forced by the relabeling rather than measured.

full rationale

FairLoop's pipeline is a human-in-the-loop fine-tuning mechanism: distill a tree from the MLP, let the user edit it, re-label the training data from the edited tree, fine-tune the MLP, and then distill a new tree to show the result. The circularity lies in the fact that the evaluation artifact (the new distilled tree) is of exactly the same kind as the intervention artifact (the edited tree), and the labels used for fine-tuning are generated by that same edited tree. Thus, if the MLP fits the relabeled data and the distillation is high-fidelity, the observed removal of a 'negatively biased' rule is a confident restatement of the relabeling rather than an independent result. The paper's explicit admission that fairness is 'assessed solely through manual inspection of the distilled decision tree' and that there are 'no built-in metrics' confirms this closed loop. No external benchmark, held-out fairness test, or user study breaks the cycle. The authors are transparent about these limitations, and the mechanism is honestly described as human-guided fine-tuning rather than a surprising empirical discovery; hence this is partial circularity (evaluation and construction collapse into one another) rather than a hidden derivation. Score 6.

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

No mathematical derivation is attempted and no fitted constants are used. The central assumptions are about distillation fidelity, label-transfer through fine-tuning, and expert judgment. The paper introduces no new physical or conceptual entities beyond the software artifact itself.

assumptions (3)
  • domain assumption The distilled decision tree faithfully approximates the MLP's behavior on relevant prefixes.
    Section 2 builds the tree from MLP predictions, but no distillation fidelity metric is reported; the tool's edits affect the MLP only through this tree.
  • domain assumption Re-labeling training data with the edited tree and fine-tuning transfers the desired behavior into the MLP without unintended side effects.
    Section 2 assumes the fine-tuned model will track the edited labels; no experiment quantifies the transfer or checks for accuracy collapse.
  • domain assumption Domain experts can reliably distinguish positive from negative bias in decision tree rules.
    Sections 1 and 3 rely on human judgment to identify negative bias; the paper states in Section 4 that no user study has been conducted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FairLoop: Software Support for Human-Centric Fairness in Predictive Business Process Monitoring." pith.science (2026). https://pith.science/paper/JOP5GV75

@misc{pith2026250820021,
  author       = {Pith},
  title        = {Pith review of: FairLoop: Software Support for Human-Centric Fairness in Predictive Business Process Monitoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JOP5GV75}},
  note         = {Machine review of arXiv:2508.20021}
}
read the original abstract

Sensitive attributes like gender or age can lead to unfair predictions in machine learning tasks such as predictive business process monitoring, particularly when used without considering context. We present FairLoop1, a tool for human-guided bias mitigation in neural network-based prediction models. FairLoop distills decision trees from neural networks, allowing users to inspect and modify unfair decision logic, which is then used to fine-tune the original model towards fairer predictions. Compared to other approaches to fairness, FairLoop enables context-aware bias removal through human involvement, addressing the influence of sensitive attributes selectively rather than excluding them uniformly.

Figures

Figures reproduced from arXiv: 2508.20021 by the authors.

Figure 1
Figure 1. Overview of the modular architecture of FairLoop. role in different activities within the process. For instance, while the gender of a patient should not influence appointment allocation, it may be essential when prescribing medication. Bias introduced by sensitive attributes that violates fairness is referred to as negative bias (e.g., using gender in appointment allocation). In contrast, bias that is necessary for… view at source ↗
Figure 2
Figure 2. Diagram describing the underlying process of the simulated cancer screening event log used for the decision tree in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Screenshot of a portion of the decision tree displayed in the frontend, for a simulated event log modeled after the cancer screening process depicted in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

  1. [11]

    Käppel, J

    M. Käppel, J. Neuberger, F. Möhrlein, S. Weinzierl, M. Matzner, S. Jablonski, A human-in-the-loop approach for improving fairness in predictive business process monitoring, in: Proceedings of the 23rd International Conference on Business Process Management, Springer, 2025. In press

    work page 2025

  2. [1]

    F. M. Maggi, C. Di Francescomarino, M. Dumas, C. Ghidini, Predictive monitoring of business processes, in: Proceedings of the 26th International Conference on Advanced Information Systems Engineering, Springer, 2014

    work page 2014

  3. [2]

    Recent Advances in Data-Driven Business Process Management

    L. Ackermann, M. Käppel, L. Marcus, L. Moder, S. Dunzer, M. Hornsteiner, A. Liessmann, Y. Zisgen, P. Empl, L.-V. Herm, et al., Recent advances in data-driven business process management, arXiv preprint arXiv:2406.01786 (2024)

    work page Pith review arXiv 2024

  4. [3]

    Di Francescomarino, C

    C. Di Francescomarino, C. Ghidini, Predictive process monitoring, in: Process Mining Handbook, Springer, 2022, pp. 320–346

    work page 2022

  5. [4]

    A. E. Márquez-Chamorro, M. Resinas, A. Ruiz-Cortés, Predictive monitoring of business processes: a survey, IEEE Transactions on Services Computing 11 (2017)

    work page 2017

  6. [5]

    de Leoni, A

    M. de Leoni, A. Padella, Achieving fairness in predictive process analytics via adversarial learning, in: CoopIS, Springer, 2024

    work page 2024

  7. [6]

    Dwork, M

    C. Dwork, M. Hardt, T. Pitassi, O. Reingold, R. Zemel, Fairness through awareness, 2011. URL: https://arxiv.org/abs/1104.3913. arXiv:1104.3913

    arXiv 2011

  8. [7]

    Oneto, S

    L. Oneto, S. Chiappa, Fairness in machine learning, in: Recent Trends in Learning From Data: Tutorials from the INNS Big Data and Deep Learning Conference, Springer, 2020

    work page 2020

Show all 15 references
  1. [8]

    Barocas, A

    S. Barocas, A. D. Selbst, Big data’s disparate impact essay, California Law Review 104 (2016) 671–732

    2016

  2. [9]

    Weinzierl, S

    S. Weinzierl, S. Zilker, S. Dunzer, M. Matzner, Machine learning in business process management: A systematic literature review, Expert Systems with Applications (2024) 124181

    2024

  3. [10]

    Käppel, L

    M. Käppel, L. Ackermann, S. Jablonski, S. Härtl, Attention please: What transformer models really learn for process prediction, in: Proceedings of the 22nd International Conference on Business Process Management, Springer Nature Switzerland, 2024, pp. 203–220

    2024

  4. [12]

    Jorbina, A

    K. Jorbina, A. Rozumnyi, I. Verenich, C. Di Francescomarino, M. Dumas, C. Ghidini, F. M. Maggi, M. La Rosa, S. Raboczi, Nirdizati: A web-based tool for predictive process monitoring, in: Proceed- ings of the BPM Demo Track and BPM Dissertation Award, co-located with 15th Inter...

    2017

  5. [13]

    M. S. Qafari, W. Van der Aalst, Fairness-aware process mining, in: On the Move to Meaningful Internet Systems: OTM 2019 Conferences: Confederated International Conferences: CoopIS, ODBASE, C&TC 2019, Springer, 2019, pp. 182–192

    2019

  6. [14]

    Peeperkorn, S

    J. Peeperkorn, S. De Vos, Achieving group fairness through independence in predictive process monitoring, arXiv preprint arXiv:2412.04914 (2024)

    2024 arXiv

  7. [15]

    de Leoni, A

    M. de Leoni, A. Padella, Achieving fairness in predictive process analytics via adversarial learning (extended version), arXiv preprint arXiv:2410.02618 (2024)

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.