Pith. sign in

REVIEW 4 major objections 5 minor 25 references

ReVise: A Human-AI Interface for Incremental Algorithmic Recourse

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ReVise turns AI recourse into an incremental, visual path.

desk verdict A promising HCI design paper for incremental recourse, but the 'demonstrates' claim outruns the evidence: twelve subjective interviews and an undefined projection heuristic. read the letter →

arxiv 2508.00002 v1 pith:U2ZCZ42I submitted 2025-07-02 cs.HC

Kaustav Bhattacharjee , Jun Yuan , Aritra Dasgupta This is my paper
classification cs.HC
keywords algorithmicrecourseincrementalvisualanalyticsSHAPattributionshuman-in-the-loopcounterfactualexplanationsinteractivevisualizationdecision-making
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 sets out to establish that algorithmic recourse should be planned incrementally—step by step, with probability feedback—rather than delivered as a single final recommendation. The authors contribute ReVise, an interactive visualization interface in which data subjects select successive recourse states, see SHAP attributions for each candidate, compare potential next steps, and build a recourse path. The paper argues this gives users greater agency and lets them make informed choices about how to improve an automated decision. Support comes from a credit-risk usage scenario and subjective feedback from twelve graduate students, with the claim that the visual encodings faithfully reflect the underlying SHAP explanations.

What carries the argument

The load-bearing mechanism is the projection, defined as the change-to-attribution ratio for a candidate next state: the magnitude of the feature change divided by the resulting SHAP attribution gain. This ratio ranks the pool of possible next targets, with the top three shown as enlarged dots, and the slope of the connected recourse path encodes the same 'minimum effort, maximum output' principle. The coordinated interface—Recourse Path Explorer and Outcome Monitor—carries this mechanism by letting users compare targets across features, see stacked positive and negative attributions, and undo poorly performing steps.

What would settle it

On a held-out set of data subjects, apply the top three projection-ranked feature changes to the model and compare the realized probability gain against changes ranked by a simple baseline such as nearest neighbor or smallest feasible change; if projection ranking does not outperform the baseline, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that presenting recourse as a sequence of user-chosen states, each guided by feature attributions and ranked by a change-to-attribution ratio, helps data subjects find recourse paths with meaningful probability gains while keeping feature changes modest. In the credit-risk scenario, the user reaches roughly 80% approval odds by iterating through states, backtracking when a step does not improve the outcome. The authors claim this workflow turns black-box model behavior into human-interpretable cues—attribution direction, deviation from averages, and trajectory slope—that support multi-way comparison of recourse paths.

Load-bearing premise

The load-bearing premise is that the projection—the change-to-attribution ratio used to rank candidate next states—actually identifies actions that improve a person's outcome, which the paper does not verify.

Editorial extensions

If this is right

  • Recourse becomes an incremental journey: users can gradually improve approval odds instead of being handed one rigid final change.
  • Data subjects can compare alternative next states and paths, backtrack from poor choices, and converge on paths with steeper attribution gains.
  • The interface grounds each step in model behavior, so choices reflect SHAP explanations rather than random trial and error.
  • The approach can transfer to other black-box decision settings with probability outputs, such as hiring or admissions, provided the same per-step data pipeline is available.

Reading between the lines

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

  • Editorial: The projection ranking is treated as a proxy for actionability, but the paper does not test whether top-projection moves are feasible or yield better real-world outcomes; a natural next step is to validate it against causal, feasibility-aware recourse baselines.
  • Editorial: Because SHAP attribution gives no guarantee of robustness, model or distribution shifts could make an already chosen path stale; the paper names recourse robustness as future work, but the current interface would not warn users of this fragility.
  • Editorial: The same visual scaffolding could support group-level recourse, where several data subjects plan coordinated changes, by extending the path explorer from one selected dot to a set of dots.
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 / 5 minor

Summary. The paper introduces ReVise, an interactive visual analytics interface for incremental algorithmic recourse. The system presents SHAP-based scatter plots of data points, an outcome monitor with stacked attributions, and trajectory visualizations, and it uses a change-to-attribution ratio (called projection) to highlight promising candidate next states. The authors report a usage scenario on the Credit Risk dataset and subjective feedback from observational studies with twelve graduate students, and they claim that the feedback demonstrates the approach can be instrumental for data subjects in choosing suitable recourse paths. The contribution is framed as shifting algorithmic recourse from final counterfactual outcomes to incremental, probability-aware, user-driven paths.

Significance. If the central claims hold, ReVise addresses a genuine gap in the algorithmic recourse literature by focusing on incremental steps and user agency rather than only on final counterfactual states. The task analysis (T1-T6) is thoughtful, the coordinated-view design is well motivated, and the system is a plausible and useful artifact for a visualization or human-AI interaction venue. The paper also has strengths in grounding the design in existing recourse and explanation literature, including acknowledging the robustness critique of attribution-based recourse. However, the significance depends on two load-bearing assumptions that are not currently established: that the projection heuristic ranks candidate next states by true recourse benefit, and that users' self-reported impressions amount to evidence of informed decision-making. The present evaluation is too thin to support the abstract's 'demonstrates' claim.

major comments (4)
  1. [Section 3.2, 'Cues for Actionability'] The change-to-attribution ratio (called projection) is never formally defined. This ratio is load-bearing because it ranks all candidate next states and enlarges the top three targets shown to users. Without a precise definition, such as the change in outcome probability per unit change in a feature, or a derivation of how SHAP values are combined across features, the reader cannot assess whether high-projection targets actually correspond to recourse states that improve the model outcome. In addition, SHAP attributions describe local model behavior at the current point, and moving to another data subject changes several features simultaneously; additive SHAP values do not by themselves imply that a summed per-feature attribution change predicts the model probability at the target state. Please provide the formula for projection and validate it against actual model outcomes at the proposed target states.
  2. [Section 5, 'Subjective User Feedback'] The claim that the results demonstrate that ReVise can be instrumental for choosing a suitable recourse path is not supported by the reported evaluation. The study involves twelve graduate students, has no baseline or comparison condition, reports no quantitative task-performance metrics, and performs no statistical analysis. The quoted themes largely concern usability and perceived usefulness, not whether users selected recourse paths with better outcomes, lower cost, or higher feasibility than alternatives. To support the central claim, the paper would need either a controlled comparison with a baseline interface or a quantitative analysis of user-selected paths (for example, improvement in model outcome, number of backtracking steps, distance to the target, and feasibility relative to automatically generated recourse paths). At minimum, the abstract and conclusion should be reworded to say the feedback provides preliminary design insights rather than demonstrating instrumental value.
  3. [Section 2 and Section 3.2] The paper cites Fokkema et al. [7], which shows that attribution-based explanations that provide recourse cannot be robust, and then asserts that limiting explanations to local regions facilitates a practical assessment of actionability. This asserted remedy is not operationalized: no locality radius is defined, no condition under which the local SHAP approximation remains valid is stated, and no validation is provided that the projection-based targets remain robust under small perturbations. Since the entire interface relies on SHAP-based cues to guide users toward better recourse paths, this is a load-bearing correctness concern rather than a minor caveat. Please specify the locality assumption and provide evidence—either theoretical or empirical—that the projected targets remain meaningful under the feature changes the interface suggests.
  4. [Section 4, 'Usage Scenario'] The usage scenario with 'Allen' is an illustrative narrative and cannot serve as evidence for the effectiveness of ReVise. The scenario demonstrates the interaction mechanics and shows that a path with substantial outcome growth was found, but it does not compare that path against alternatives, quantify the cost of the changes, or check whether the selected path is actually feasible or optimal. Please present the scenario as a demonstration of functionality, not as validation of the interface's decision-support value.
minor comments (5)
  1. [Section 3.2, 'Cues for Actionability'] The paper does not explain why exactly three target candidates are highlighted or why exactly six features are shown in the Recourse Path Explorer; a brief rationale or a note on the sensitivity of these choices would improve replicability.
  2. [Figure 1 and Section 4] The figure is referenced as Figure 1a-1d but is not included in the text of the manuscript; when the final version includes the figure, please ensure all panels are legible and that each panel referenced in the usage scenario is clearly labeled.
  3. [Section 5, 'Subjective User Feedback'] One participant is quoted as saying the results are based on 'real students' data,' but the paper uses the Credit Risk dataset from Kaggle; please clarify whether the study used a different dataset or correct the quote so that it does not mislead readers.
  4. [Section 4 and Figure 1 caption] The statement that ReVise's visual encodings 'faithfully reflect the explanation results offered by SHAP' is stronger than what the study measures, which is self-reported consistency; please soften this phrasing to 'are consistent with' or similar.
  5. [Section 3.2, notation] The term 'projection' is used informally; please introduce a notation for the change-to-attribution ratio and define it explicitly in a table or equation so that the reader can connect the described visual cue to a computable quantity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ReVise is a visualization system paper whose claims rest on interface construction, external SHAP values, and subjective feedback, not on a derivation that reduces to its own inputs.

full rationale

The paper contains no fitted parameters, derived equations, or uniqueness theorems whose conclusions are forced by their own definitions. The central artifact is an interactive interface that displays SHAP attributions computed by an external method (Lundberg and Lee), and the reported evaluation is qualitative feedback from twelve graduate students. The projection heuristic in Section 3.2 ('Cues for Actionability') is an unvalidated design choice, which is a correctness risk, not a circularity: the paper does not claim to derive the top-target ranking from a definitional identity or from a prior self-citation. The only quasi-circular element is the faithfulness check in Section 5, where users are asked whether the visual encodings match the SHAP values from which those encodings were built; that is a design-consistency check by construction and is not presented as an independent confirmation of an external prediction. The one self-citation (ViCE, reference [9], which shares an author) is used as related work describing an existing interface and carries no load-bearing argument. Accordingly, no circular step meets the evidentiary bar of quote-and-reduction, and the honest finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper's central claim rests on the validity of SHAP-based guidance, the representativeness of the public credit dataset, a small qualitative user study, and a hand-chosen projection heuristic; none of these are independently verified in the paper.

free parameters (2)
  • Number of highlighted target candidates = 3
    Selected by hand in Section 3.2; based on qualitative feedback, not on measured recourse success.
  • Number of features in Recourse Path Explorer = 6
    Chosen to reduce cognitive load in Section 3.3.1 and 3.3.2; arbitrary, no sensitivity analysis.
assumptions (5)
  • domain assumption SHAP attributions are a valid local explanation of model behavior and a suitable guide for recourse actions.
    Used throughout Section 3.2; the paper cites Fokkema et al. [7] questioning robustness but does not resolve the tension.
  • domain assumption The Credit Risk dataset and the SHAP-based model used in the scenario are representative of realistic recourse settings.
    Section 4 usage scenario depends on this; no generalizability analysis.
  • domain assumption Subjective feedback from twelve graduate students is sufficient evidence for the claim that the interface is instrumental for data subjects.
    Section 5 is the sole empirical evaluation; the sample is small, self-selected, and computationally trained.
  • ad hoc to paper The change-to-attribution ratio ranks candidate next states by recourse benefit.
    Section 3.2 defines this heuristic without external validation.
  • domain assumption Incremental recourse is more beneficial to users than single-step recourse.
    Motivates the paper in Section 1 but is not explicitly tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReVise: A Human-AI Interface for Incremental Algorithmic Recourse." pith.science (2026). https://pith.science/paper/U2ZCZ42I

@misc{pith2026250800002,
  author       = {Pith},
  title        = {Pith review of: ReVise: A Human-AI Interface for Incremental Algorithmic Recourse},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U2ZCZ42I}},
  note         = {Machine review of arXiv:2508.00002}
}
read the original abstract

The recent adoption of artificial intelligence in socio-technical systems raises concerns about the black-box nature of the resulting decisions in fields such as hiring, finance, admissions, etc. If data subjects -- such as job applicants, loan applicants, and students -- receive an unfavorable outcome, they may be interested in algorithmic recourse, which involves updating certain features to yield a more favorable result when re-evaluated by algorithmic decision-making. Unfortunately, when individuals do not fully understand the incremental steps needed to change their circumstances, they risk following misguided paths that can lead to significant, long-term adverse consequences. Existing recourse approaches focus exclusively on the final recourse goal but neglect the possible incremental steps to reach the goal with real-life constraints, user preferences, and model artifacts. To address this gap, we formulate a visual analytic workflow for incremental recourse planning in collaboration with AI/ML experts and contribute an interactive visualization interface that helps data subjects efficiently navigate the recourse alternatives and make an informed decision. We present a usage scenario and subjective feedback from observational studies with twelve graduate students using a real-world dataset, which demonstrates that our approach can be instrumental for data subjects in choosing a suitable recourse path.

Figures

Figures reproduced from arXiv: 2508.00002 by the authors.

Figure 1
Figure 1. Recourse Planning Usage Scenario: (a) Utilizing the Credit Risk dataset, the user selected a recourse path from the Recourse Path Explorer. (b) After multiple iterations, she identified a recourse path that resulted in significant improvement in the model outcome. (c) At the same time, the user observed that some features reached significantly higher-than-average values, while others showed a downward trend in their… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 16 canonical work pages

  1. [7]

    Fokkema, R

    H. Fokkema, R. de Heide, and T. van Erven. Attribution-based Expla- nations that Provide Recourse Cannot be Robust. Journal of Machine Learning Research, 24(360):1–37, 2023. 2

    work page 2023

  2. [1]

    Albini, J

    E. Albini, J. Long, D. Dervovic, and D. Magazzeni. Counterfactual Shapley Additive Explanations. In 2022 ACM Conference on Fair- ness, Accountability, and Transparency, pp. 1054–1070. ACM, Seoul Republic of Korea, June 2022. doi: 10.1145/3531146.3533168 2

    arXiv 2022

  3. [2]

    Cheng, Y

    F. Cheng, Y . Ming, and H. Qu. DECE: Decision explorer with counter- factual explanations for machine learning models. IEEE Transactions on Visualization and Computer Graphics, 27(2):1438–1447, 2020. 2

    work page 2020

  4. [3]

    https://www.kaggle.com/datasets/ laotse/credit-risk-dataset/data

    Credit Risk Dataset. https://www.kaggle.com/datasets/ laotse/credit-risk-dataset/data . Accessed: 2024-03-31. 4

    work page 2024

  5. [4]

    Dimara and J

    E. Dimara and J. Stasko. A Critical Reflection on Visualization Re- search: Where Do Decision Making Tasks Hide? IEEE Transac- tions on Visualization and Computer Graphics, 28(1):1128–1138, Jan

    work page

  6. [5]

    Fern and Q

    X. Fern and Q. Pope. Text Counterfactuals via Latent Optimization and Shapley-Guided Search. In M.-F. Moens, X. Huang, L. Spe- cia, and S. W.-t. Yih, eds., Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 5578–5593. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic, Nov. 2021. d...

    doi:10.18653/v1/2021.emnlp-main 2021

  7. [6]

    Fern´andez-Lor´ıa, F

    C. Fern´andez-Lor´ıa, F. Provost, and X. Han. Explaining Data-Driven Decisions made by AI Systems: The Counterfactual Approach, Oct

    work page

  8. [8]

    Fonseca, A

    J. Fonseca, A. Bell, C. Abrate, F. Bonchi, and J. Stoyanovich. Setting the right expectations: Algorithmic recourse over time. In Proceed- ings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, pp. 1–11, 2023. 4

    work page 2023

Show all 25 references
  1. [9]

    Gomez, S

    O. Gomez, S. Holter, J. Yuan, and E. Bertini. ViCE: Visual coun- terfactual explanations for machine learning models. In Proceedings of the 25th International Conference on Intelligent User Interfaces , pp. 531–535. ACM, Cagliari Italy, Mar. 2020. doi: 10.1145/3377325. 3377536 2

    2020 doi

  2. [10]

    R. Guo, L. Cheng, J. Li, P. R. Hahn, and H. Liu. A survey of learning causality with data: Problems and methods. ACM Computing Surveys (CSUR), 53(4):1–37, 2020. 1

    2020

  3. [11]

    Karimi, G

    A.-H. Karimi, G. Barthe, B. Sch ¨olkopf, and I. Valera. A Survey of Algorithmic Recourse: Contrastive Explanations and Consequential Recommendations. ACM Computing Surveys, 55(5):1–29, May 2023. doi: 10.1145/3527848 1

    2023 doi

  4. [12]

    Karimi, B

    A.-H. Karimi, B. Sch ¨olkopf, and I. Valera. Algorithmic Recourse: From Counterfactual Explanations to Interventions. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Trans- parency, pp. 353–362. ACM, Virtual Event Canada, Mar. 2021. doi: 10.1145/3442...

    2021

  5. [13]

    S. Koh, B. H. Kim, and S. Jo. Understanding the user perception and experience of interactive algorithmic recourse customization. ACM Transactions on Computer-Human Interaction, 31(3):1–25, 2025. 4

    2025

  6. [14]

    Kommiya Mothilal, D

    R. Kommiya Mothilal, D. Mahajan, C. Tan, and A. Sharma. Towards Unifying Feature Attribution and Counterfactual Explanations: Dif- ferent Means to the Same End. InProceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society , pp. 652–663. ACM, Virtual Event USA, July...

    2021

  7. [15]

    H. Lam, E. Bertini, P. Isenberg, C. Plaisant, and S. Carpendale. Empir- ical studies in information visualization: Seven scenarios.IEEE trans- actions on visualization and computer graphics , 18(9):1520–1536,

  8. [16]

    S. M. Lundberg and S.-I. Lee. A unified approach to interpreting model predictions. Advances in neural information processing sys- tems, 30, 2017. 2

    2017

  9. [17]

    T. Miller. Contrastive Explanation: A Structural-Model Approach. The Knowledge Engineering Review , 36:e14, 2021. doi: 10.1017/ S0269888921000102 1

    2021

  10. [18]

    Poyiadzi, K

    R. Poyiadzi, K. Sokol, R. Santos-Rodriguez, T. De Bie, and P. Flach. FACE: feasible and actionable counterfactual explanations. In Pro- ceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pp. 344–350, 2020. 1

    2020

  11. [19]

    Ramon, D

    Y . Ramon, D. Martens, F. Provost, and T. Evgeniou. A compari- son of instance-level counterfactual explanation algorithms for behav- ioral and textual data: SEDC, LIME-C and SHAP-C. Advances in Data Analysis and Classification , Sept. 2020. doi: 10.1007/s11634 -020-00418-3 2

    2020 doi

  12. [20]

    S. Rathi. Generating Counterfactual and Contrastive Explanations us- ing SHAP, June 2019. 2

    2019

  13. [21]

    Upadhyay, S

    S. Upadhyay, S. Joshi, and H. Lakkaraju. Towards Robust and Reli- able Algorithmic Recourse. In Advances in Neural Information Pro- cessing Systems, vol. 34, pp. 16926–16937. Curran Associates, Inc.,

  14. [22]

    Ustun, A

    B. Ustun, A. Spangher, and Y . Liu. Actionable Recourse in Linear Classification. In Proceedings of the Conference on Fairness, Ac- countability, and Transparency, pp. 10–19. ACM, Atlanta GA USA, Jan. 2019. doi: 10.1145/3287560.3287566 1

    2019

  15. [23]

    Z. J. Wang, J. Wortman Vaughan, R. Caruana, and D. H. Chau. GAM Coach: Towards interactive and user-centered algorithmic recourse. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pp. 1–20, 2023. 2

    2023

  16. [24]

    Wexler, M

    J. Wexler, M. Pushkarna, T. Bolukbasi, M. Wattenberg, F. Viegas, and J. Wilson. The What-If Tool: Interactive Probing of Machine Learning Models. IEEE Transactions on Visualization and Computer Graphics, pp. 1–1, 2019. doi: 10.1109/TVCG.2019.2934619 2

    2019

  17. [2022]

    doi: 10.1109/TVCG.2021.3114813 1, 2

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.