Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Robust Counterfactual Explanations under Model Multiplicity Using Multi-Objective Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper proposes generating counterfactual explanations as Pareto-improving solutions over a set of equally accurate models, and reports that the resulting explanations are more robust than single-model or stacked baselines.

desk verdict The idea of using Pareto improvement for counterfactual explanations under model multiplicity is reasonable, but the paper's core objective is undefined, the empirical evidence is missing from the text, and the stated method cannot be reproduced. read the letter →

arxiv 2501.05795 v4 pith:AT5R54BS submitted 2025-01-10 cs.LG cs.AI

Keita Kinjo This is my paper · ORCID
classification cs.LGcs.AI
keywords counterfactualexplanationsalgorithmicrecoursemodelmultiplicityParetoimprovementmulti-objectiveoptimizationNSGA-IIexplainableartificialintelligencerobustness
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

The paper claims that a counterfactual explanation can be made robust to model multiplicity by requiring it to Pareto-improve a set of similarly accurate models rather than any single model. It proposes generating such explanations by multi-objective optimization over the vector of prediction losses, one loss per model, and evaluates them with a new True Improvement Ratio that checks how often the explanation improves on the true data-generating function. Experiments on simulated and real data lead the author to conclude that the method is both robust and practical: it achieves higher validity relative to dissimilarity and plausibility, works for regression and classification, allows constraints on the explanation, and produces a diverse set of candidate explanations. The motivation is that in automated machine-learning settings many models can have the same accuracy, and safe decisions in domains such as lending, marketing, and medical care should not depend on which of those models happens to be chosen.

What carries the argument

The machinery is multi-objective optimization over the vector of model losses, solved with NSGA-II (a fast elitist non-dominated sorting genetic algorithm) implemented through an evolutionary computation library. The vector objective is $F(X_{cf})=(\mathrm{loss}(y_t,f_1(X_{cf})),\ldots,\mathrm{loss}(y_t,f_m(X_{cf})))$, constrained by $d(X_b,X_{cf})\le C$ and user-supplied equality and inequality constraints; the Pareto set gives diverse candidate explanations. The new evaluation index, True Improvement Ratio (TIR), computes the fraction of generated CEs for which the true function's value improves over the base prediction, and the paper uses average validity-over-dissimilarity and validity-over-plausibility ratios to compare methods.

What would settle it

Take a dataset with a known true function, generate CEs with Method 3 using the top-m models, then compute the True Improvement Ratio on a different model with equal test accuracy that was not among the m; if the held-out TIR is no better than single-model or stacked baselines, the robustness claim over model multiplicity is contradicted.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that Pareto improvement supplies the missing link between counterfactual explanation and model multiplicity. Given $m$ models $f_1,\ldots,f_m$ with similar test accuracy, the paper defines a robust CE as a solution $X_{cf}^*$ minimizing the vector $(\mathrm{loss}(y_t,f_1(X_{cf})),\ldots,\mathrm{loss}(y_t,f_m(X_{cf})))$ subject to a distance bound $d(X_b,X_{cf})\le C$ and arbitrary constraints. Because the minimization is multi-objective, the output is a Pareto set rather than a single point, so the explanation improves at least one model's loss without worsening the others. The paper reports that using the top two or three models by accuracy and searching with NSGA-II gives consistently higher True Improvement Ratios than single-model baselines and stacked-model weighting, at lower dissimilarity and plausibility cost, and that the same procedure transfers to a real dataset with intervention-type features.

Load-bearing premise

The load-bearing premise is that the top-m models selected by test accuracy stand in for all equally accurate models, so a counterfactual that Pareto-improves them is assumed to also improve the true outcome and any other model outside the chosen set, yet the Section 2.4 True Improvement Ratio checks improvement only for those selected models and no held-out model validation is provided.

Editorial extensions

If this is right

  • A counterfactual that Pareto-improves a set of equally accurate models can be found for regression as well as classification, since the loss vector is defined for continuous targets.
  • User-specified constraints, such as sign restrictions on features or a budget on the amount of change, can be embedded directly into the optimization, so the robust CE respects real-world feasibility.
  • Generating a Pareto set instead of a single CE lets a decision maker pick the most acceptable explanation, or use the medoid or centroid as a safe default.
  • In simulated settings where the true function is known, the Pareto-based method yields a higher fraction of genuinely improving CEs (TIR) than single-model or stacked baselines.
  • The approach extends to automated machine-learning settings where many models have comparable accuracy, reducing the dependence of recourse on an arbitrary model choice.

Reading between the lines

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

  • Editorial inference: if Pareto improvement over the top-m models transfers to other equally accurate models, the method could serve as a model-selection-free recourse procedure in automated machine-learning pipelines; the paper does not test held-out models, since the Section 2.4 True Improvement Ratio is computed only for the selected models and a known true function.
  • Editorial inference: the Pareto-front framing connects CE robustness to social choice; utilitarian selection (maximize average gain) and Rawlsian selection (maximize the worst-case gain) could be tested as preference-based extensions.
  • Editorial inference: because NSGA-II is stochastic, the Pareto set varies across runs; measuring the stability of the selected medoid CE across random seeds would quantify how much of the reported robustness comes from the method rather than a particular run.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 proposes a method for generating counterfactual explanations (CEs) that are robust under model multiplicity. It formulates the problem as a multi-objective optimization over m models with similar accuracy, defines the desired solution as a Pareto improvement, and proposes to solve it with NSGA-II. The experimental sections compare this method (Method 3) with single-model optimization (Method 1) and a stacking-based weighted objective (Method 2), using simulated data with a known true function and real survey data on academic achievement. The central claim is that Method 3 produces CEs that maintain higher validity/dissimilarity and validity/plausibility ratios, achieve higher True Improvement Ratio, and are therefore both robust and practical.

Significance. The conceptual direction is relevant: applying Pareto improvement to CE robustness under model multiplicity, and extending the setting to regression with constraints, addresses a real gap in the literature. The proposed True Improvement Ratio is also a reasonable simulation-only evaluation idea, since it checks the CE against an independently specified true function. However, the central method as written is not well defined because the objective is stated with an infinite target value, and the empirical evidence is incomplete because several key tables and figures are missing from the manuscript. If the formulation and experiments were corrected, the idea could be worth further study, but in its current form the paper does not support its claims.

major comments (4)
  1. [§2.1 and §3.1] The optimization problem in §2.1 is not well defined for the experimental setting. The text states that for continuous y the loss is squared or absolute error between yt and f_j(Xcf), and §3.1 then sets yt = ∞. With yt = ∞, loss(yt, f_j(Xcf)) is infinite for every finite prediction, so all m objectives are identical and constant over the feasible set; no NSGA-II run on these objectives can select one counterfactual over another. The later statement in §3.1 that “this study simply used the predicted values of y, where higher values indicated better evaluations” describes a different objective, such as minimizing -f_j(Xcf), which is never formally defined. As a result, Method 3 is not reproducible from the manuscript, and the reported FIR and balance ratios are not tied to the formalized problem. This is a load-bearing gap because it affects the definition of the proposed method and every experiment that uses it.
  2. [§2.1 and §2.2] Pareto optimality of X_cf with respect to the vector of losses does not imply that X_cf is a Pareto improvement over the base point X_b. A Pareto-optimal solution can have higher loss than f_j(X_b) in every objective when the base point is not in the feasible set, and in general the argmin of the objective vector need not dominate the base input. The paper's phrase “improves consistently in all models” requires a condition such as f_j(X_cf) > f_j(X_b) for all j, or loss(yt, f_j(Xcf)) < loss(yt, f_j(Xb)); no such constraint or post-hoc check is present in §2.1. The experiments should verify per-model improvement over the base, and the formulation should include this condition explicitly.
  3. [§2.3 and §2.4] The selection of the top-m models by test accuracy in Process 3 makes the robustness claim contingent on the representativeness of that particular set, but the paper provides no justification or bound. The True Improvement Ratio in §2.4 is evaluated against the true function only for the same selected models used to generate the CEs; therefore a high TIR shows improvement relative to those selected models and to the true function at those inputs, not robustness to arbitrary equally accurate models in the multiplicity set. A concrete fix would be to hold out several accurate models not used in generating the CEs and report FIR on them, or to provide a finite-sample or Lipschitz-style bound over the set of models within a given accuracy threshold. This is central because the paper's claim is robustness under model multiplicity, not only under the chosen top-m models.
  4. [§3.1, §3.2, and Tables 3–8] The experimental evidence as presented is incomplete. In the provided manuscript, Tables 3, 4, 5, 7, and 8 contain only captions and no data, and Figures 1–4 appear as captions only, so the reported comparisons of val, dissim, plaus, FIR, and the ratio values cannot be checked. In addition, Table 6 in §3.2 reproduces the same numerical values as Table 1 and includes Model 3 and the stacking model, while the text says the real-data analysis uses Model 1, Model 2, and Model 4; this suggests a copy-and-paste error and makes the real-data experiment unverifiable. These are not cosmetic issues because the central claim that Method 3 could be confirmed as robust rests on those tables.
minor comments (5)
  1. [§2.4] In the definition of Val_j, the notation |yt - f_j(X*)| is used with yt = ∞, which is inconsistent with the sentence explaining that higher predicted values are better; please define Val as an explicit function of f_j(X*) when maximization is intended, and define the corresponding objective in §2.1.
  2. [§2.2] The notation min_{θ ∈ C} F(θ) uses C for a constraint set, but C was previously used for the distance bound in §2.1; the constraint set is also not defined in §2.2. Please use a distinct symbol and define the feasible set precisely.
  3. [§2.4] In the TIR formula, ys is not defined; if it denotes the true function value at the base point, this should be stated explicitly, and the comparison should be tied to f_j(X_b) for each model if the claim is per-model improvement.
  4. [§3.2] The sentence “we set the condition that the improvement is negative when the value is 1 and positive when the value is 0 in T1–T19” is unclear and does not correspond to any constraint form introduced in §2.1; please specify the constraint mathematically.
  5. [General] No implementation details for NSGA-II are given (population size, generations, crossover and mutation rates, seed), and no code or data are provided; given that evolutionary computation is stochastic, these details are necessary for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central robustness claim is checked against an independently specified true function in simulation, and no fitted parameter is fed back into the definition of the result.

full rationale

The paper's derivation chain is self-contained rather than circular. The proposed method in Section 2.3 optimizes the vector of losses over the top-m accurate models to obtain Pareto-optimal counterfactual explanations; the evaluation in Section 2.4 then measures validity, dissimilarity, plausibility, and a True Improvement Ratio. The load-bearing robustness claim in Section 3.1 ('Method 3 ... achieved consistently higher FIR, and thus could be confirmed as robust') is not equivalent to the optimization objective: TIR/FIR is computed from the true simulation function ('this can only be used when the true function is known'), which is external to the models whose predictions were optimized. No parameter is fitted to the evaluation data and then renamed as a prediction; C and lambda are fixed constants. The only mild concern is that the top-m model selection uses the test split, but this affects representativeness rather than creating a definitional identity between input and output. The noted issue that setting yt = Infinity makes the formal loss undefined in Section 2.1, with Section 3.1 instead using raw predicted values, is a correctness/reproducibility gap, not a circular reduction: it does not make the reported result equal to its input by construction. There are no load-bearing self-citations or imported uniqueness theorems. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-chosen optimization bounds, the number of selected models, and unstated NSGA-II settings. It also depends on the assumption that a small set of high-accuracy models represents the full model multiplicity, and on a causal interpretation of feature changes. No new entities are introduced beyond the proposed TIR metric, which is a measurement definition rather than an entity.

free parameters (5)
  • Distance bound C = 3 (simulation), 5 (real data)
    Hand-chosen constraint on how far CEs can move from the base point; directly controls the feasible region and the trade-off between cost and improvement (Sections 3.1 and 3.2).
  • Regularization weight lambda = 2
    Hand-chosen weight in the baseline single-model objectives balancing prediction loss against distance; affects the CE comparison (Section 3.1).
  • Number of selected models m = 2, 3, 4 in simulation; 3 in real data
    The robust CE is defined only over the top-m models; choosing m is a modeling decision that changes the Pareto front (Sections 2.3 and 3.1).
  • NSGA-II hyperparameters = not reported
    Population size, generations, crossover and mutation rates are not stated; the quality and diversity of the Pareto front depend on them (Section 2.3).
  • Evaluation sample sizes = 20 CEs per base; 50 base cases
    Chosen by hand with no power analysis; affects the stability of all reported averages (Section 3.1).
assumptions (4)
  • standard math Pareto optimality and NSGA-II approximate behavior are standard and correctly implemented.
    Section 2.2 uses standard definitions; the paper itself notes that evolutionary computation gives no guarantee of finding a true Pareto solution.
  • domain assumption The top-m models by test accuracy adequately represent the model multiplicity and the true function.
    Section 2.3 Process 3 selects models by accuracy; the robustness claim relies on this representative set, but no held-out model validation or theory is provided.
  • domain assumption Changing the features in a CE will cause the desired outcome in the real world.
    Section 3.2 interprets features T1-T19 as interventions and calls them 'intervention rate', but the paper does not establish causal validity.
  • ad hoc to paper Optimizing with an infinite target value is a well-posed way to encode 'increase the prediction'.
    Section 2.1 defines loss with a finite yt; Section 2.4 sets yt=infinity, which makes the squared or absolute loss undefined, and then silently switches to maximizing predictions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Counterfactual Explanations under Model Multiplicity Using Multi-Objective Optimization." pith.science (2026). https://pith.science/paper/AT5R54BS

@misc{pith2026250105795,
  author       = {Pith},
  title        = {Pith review of: Robust Counterfactual Explanations under Model Multiplicity Using Multi-Objective Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AT5R54BS}},
  note         = {Machine review of arXiv:2501.05795}
}
read the original abstract

In recent years, explainability in machine learning has gained importance. In this context, counterfactual explanation (CE), which is an explanation method that uses examples, has attracted attention. However, it has been pointed out that CE is not robust when there are multiple machine-learning models with similar accuracy. These problems are important when using machine learning to make safe decisions. In this paper, we propose robust CEs that introduce a new viewpoint -- Pareto improvement -- and a method that uses multi-objective optimization to generate it. To evaluate the proposed method, we conducted experiments using both simulated and real data. The results demonstrate that the proposed method is both robust and practical. This study highlights the potential of ensuring robustness in decision-making by applying the concept of social welfare. We believe that this research can serve as a valuable foundation for various fields, including explainability in machine learning, decision-making, and action planning based on machine learning.

Figures

Figures reproduced from arXiv: 2501.05795 by the authors.

Figure 1
Figure 1. Average of averages of CEs, including base data [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Average of averages of CEs, excluding base data [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Pareto front(red: base data; blue: CEs) 14 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Average of CEs (vertical axis: average; error bars: standard deviation) Therefore, it is important to improve T2, T4, T8, T12, T13, T14, T18, and so on. In other words, Knowledge2, Knowledge4, Recording study time, Group leader￾ship experience, Opinion presentation and…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Generalized-Bayes Perspective on Counterfactual Explanations: Posterior-Based Decision-Making and Evaluation

    cs.AI 2026-07 conditional novelty 5.0 of 10

    Counterfactual explanation by distance minimization is re-derived as MAP inference in a generalized Bayes posterior, and additional posterior-based decision rules are proposed and evaluated.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    P. P. Angelov, E. A. Soares, R. Jiang, N. I. Arnold, and P. M. Atk inson. Explainable artificial intelligence: An analytical review. Wiley Interdis- ciplinary Reviews: Data Mining and Knowledge Discovery , 11(5):e1424, 2021

    work page 2021

  2. [2]

    Blank and K

    J. Blank and K. Deb. Pymoo: Multi-objective optimization in python . IEEE Access, 8:89497–89509, 2020

    work page 2020

  3. [3]

    Dandl, C

    S. Dandl, C. Molnar, M. Binder, and B. Bischl. Multi-objective coun terfac- tual explanations. In International Conference on Parallel Problem Solving from Nature, pages 448–469. Springer International Publishing, 2020

    work page 2020

  4. [4]

    K. Deb. Multi-objective optimisation using evolutionary algorithms : An introduction. In Multi-objective Evolutionary Optimisation for Product De - sign and Manufacturing , pages 3–34. Springer, 2011

    work page 2011

  5. [5]

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitis t multi- objective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2):182–197, 2002

    work page 2002

  6. [6]

    K. Deb, K. Sindhya, and J. Hakanen. Multi-objective optimization . In Decision Sciences, pages 161–200. CRC Press, 2016

    work page 2016

  7. [7]

    Dwivedi, D

    R. Dwivedi, D. Dave, H. Naik, S. Singhal, R. Omer, P. Patel, and R. R an- jan. Explainable AI (XAI): Core ideas, techniques, and solutions. ACM Computing Surveys , 55(9):1–33, 2023

    work page 2023

  8. [8]

    Fliege and B

    J. Fliege and B. F. Svaiter. Steepest descent methods for multic riteria optimization. Mathematical Methods of Operations Research , 51(3):479– 494, 2000

    work page 2000

Show all 27 references
  1. [9]

    Freiesleben

    T. Freiesleben. The intriguing relation between counterfactual explanations and adversarial examples. Minds and Machines , 32(1):77–109, 2022

    2022

  2. [10]

    Guidotti

    R. Guidotti. Counterfactual explanations and how to find them : Litera- ture review and benchmarking. Data Mining and Knowledge Discovery , 38(5):2770–2824, 2024

    2024

  3. [11]

    Gunantara

    N. Gunantara. A review of multi-objective optimization: Method s and its applications. Cogent Engineering, 5(1):1502242, 2018

    2018

  4. [12]

    Jiang, F

    J. Jiang, F. Leofante, A. Rago, and F. Toni. Robust counterf actual expla- nations in machine learning: A survey. arXiv preprint arXiv:2402.01928 , 2024

    2024 arXiv

  5. [13]

    Jiang, A

    J. Jiang, A. Rago, F. Leofante, and F. Toni. Recourse under m odel multi- plicity via argumentative ensembling. Technical report, CoRR, 2023 . 17

    2023

  6. [14]

    A. H. Karimi, G. Barthe, B. Sch¨ olkopf, and I. Valera. A survey of algo- rithmic recourse: Contrastive explanations and consequential re commen- dations. ACM Computing Surveys , 55(5):1–29, 2022

    2022

  7. [15]

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T. Liu. LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems , volume 30, 2017

    2017

  8. [16]

    Krishna, J

    S. Krishna, J. Ma, and H. Lakkaraju. Towards bridging the gap s between the right to explanation and the right to be forgotten. In International Conference on Machine Learning , pages 17808–17826. PMLR, 2023

    2023

  9. [17]

    Leofante, E

    F. Leofante, E. Botoeva, and V. Rajani. Counterfactual ex planations and model multiplicity: A relational verification view. In Proceedings of the International Conference on Principles of Knowledge Repre sentation and Reasoning, volume 19, pages 763–768, 2023

    2023

  10. [18]

    Linardatos, V

    P. Linardatos, V. Papastefanopoulos, and S. Kotsiantis. Exp lainable AI: A review of machine learning interpretability methods. Entropy, 23(1):18, 2020

    2020

  11. [19]

    Mercier, F

    Q. Mercier, F. Poirion, and J. A. D´ esid´ eri. A stochastic multiple gradient descent algorithm. European Journal of Operational Research , 271(3):808– 817, 2018

    2018

  12. [20]

    V. Pareto. Manual of Political Economy: A Critical and Variorum Editio n. Oxford University Press, 2014

    2014

  13. [21]

    Pawelczyk, C

    M. Pawelczyk, C. Agarwal, S. Joshi, S. Upadhyay, and H. Lakka raju. Ex- ploring counterfactual explanations through the lens of adversa rial exam- ples: A theoretical and empirical analysis. In International Conference on Artificial Intelligence and Statistics , pages 4574–4594...

    2022

  14. [22]

    Pawelczyk, K

    M. Pawelczyk, K. Broelemann, and G. Kasneci. On counterfact ual ex- planations under predictive multiplicity. In Conference on Uncertainty in Artificial Intelligence , pages 809–818. PMLR, 2020

    2020

  15. [23]

    Pawelczyk, T

    M. Pawelczyk, T. Datta, J. van-den Heuvel, G. Kasneci, and H. Lakkaraju. Probabilistically robust recourse: Navigating the trade-offs betwe en costs and robustness in algorithmic recourse. In 11th International Conference on Learning Representations , 2023

    2023

  16. [24]

    M. J. Powell. A direct search optimization method that models the ob- jective and constraint functions by linear interpolation. In Operations Re- search and Engineering , pages 51–67. Springer, 1994

    1994

  17. [25]

    Slack, A

    D. Slack, A. Hilgard, H. Lakkaraju, and S. Singh. Counterfact ual expla- nations can be manipulated. Advances in Neural Information Processing Systems, 34:62–75, 2021. 18

    2021

  18. [26]

    Upadhyay, S

    S. Upadhyay, S. Joshi, and H. Lakkaraju. Towards robust an d reliable algorithmic recourse. Advances in Neural Information Processing Systems , 34:16926–16937, 2021

    2021

  19. [27]

    Verma, V

    S. Verma, V. Boonsanong, M. Hoang, K. Hines, J. Dickerson, a nd C. Shah. Counterfactual explanations and algorithmic recourses for mach ine learn- ing: A review. ACM Computing Surveys , 56(12):1–42, 2024. Appendix: Types of Educational Interventions Experienced by Each Individ...

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.