Pith. sign in

REVIEW 3 major objections 5 minor 16 references

This paper proves that the standard distance-minimization counterfactual explanation is exactly the maximum a posteriori estimate of a Gibbs posterior built from a distance-based prior, and uses that equivalence to derive new decision rules

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 14:02 UTC pith:IOXZ3FD7

load-bearing objection The MAP–Gibbs equivalence is a definitional identity, not a deep theoretical result; the paper's real value is the CVaR-CE and model-mixture extensions, though the empirical support is thin. the 3 major comments →

arxiv 2607.29077 v1 pith:IOXZ3FD7 submitted 2026-07-31 cs.AI stat.ML

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

Keita Kinjo This is my paper
classification cs.AI stat.ML
keywords counterfactual explanationsgeneralized BayesGibbs posteriorMAP estimationmodel multiplicityCVaRalgorithmic recourseBayesian decision theory
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that the common practice of finding a counterfactual by minimizing a weighted sum of goal-attainment loss and change distance is not just a heuristic: it is exactly MAP estimation of a Gibbs posterior when the prior is chosen as the exponential of the distance. If true, every existing distance-minimizing counterfactual method inherits a probabilistic interpretation, and the full posterior over counterfactual candidates becomes available for decision-making. The paper builds on this posterior to define three decision rules—MAP, posterior-mean Bayes, and risk-averse CVaR—along with distribution-level metrics such as success probability, tail loss, stability, and variable importance. A reader should care because this recasts a widely used explainability tool as a Bayesian inference problem, enabling model averaging across multiple competing models and principled handling of recourse uncertainty.

Core claim

The central discovery is Proposition 1: the minimizer of ηℓ(f(˜x), y*) + d(xb, ˜x) is exactly the maximizer of the Gibbs posterior p(˜x|xb, y*) ∝ exp(−ηℓ(f(˜x), y*)) exp(−d(xb, ˜x)), formed by exponentiating the goal-attainment loss and using the distance as a prior. The paper calls this the Distance-Prior Generalized Bayes CE (DP-GBCE). This is an exact algebraic equivalence, not a numerical approximation, and it holds for any loss and any distance that make the posterior integrable. On this basis the paper argues that distance-minimization counterfactual explanations are a special case of probabilistic inference, and that the entire posterior distribution—not just its mode—is a legitimate

What carries the argument

The load-bearing object is the Gibbs posterior over counterfactual candidates, p(˜x|xb, y*) ∝ exp(−ηℓ(f(˜x), y*)) exp(−d(xb, ˜x)), with the exponential of the change distance serving as the prior. Proposition 1 shows that maximizing this posterior is identical to minimizing the conventional counterfactual objective. In experiments, the posterior is approximated by importance sampling from a Gaussian proposal centered at the base point, after which decision rules (MAP via density estimation, posterior mean, CVaR-CE) and distribution-level metrics are applied. The model-uncertainty extension, ModelUnc, replaces the single loss with a weighted mixture of exponentials from multiple models, refle

Load-bearing premise

The entire equivalence rests on choosing the prior over counterfactuals to be exactly the exponential of the change distance; if that prior is not a faithful representation of what counts as a feasible or natural change, the posterior and the decision rules inherit that arbitrary choice.

What would settle it

Run the paper's importance-sampling MAP approximation side by side with direct multistart optimization of the same objective on many base points. If the sample-based MAP is consistently far from the direct optimum (as the paper's own 10D results already suggest, e.g., Gibbs MAP loss 1.816 versus direct optimization loss 0.017), then the practical claim that the method recovers the posterior mode is not supported; the theoretical equivalence itself would remain intact, but its algorithmic realization would be falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Every existing distance-minimizing counterfactual method is, in effect, computing a MAP estimate of a Gibbs posterior; the equivalence provides a probabilistic justification for the cost-minimization formulation.
  • Replacing the point estimate with the full posterior enables two additional decision rules—Bayes decision (posterior mean or geometric median) and CVaR-CE—which trade off goal attainment, change distance, and execution robustness differently.
  • Using a mixture of model-specific posteriors with Bayesian weights gives a distributional counterfactual that explicitly accounts for model multiplicity, and the paper shows that this mixture increases tail risk and posterior dispersion compared with single-model posteriors.
  • Distribution-level metrics (success probability, tail loss, stability, and distributional variable importance) make it possible to evaluate the whole counterfactual posterior, not just one explanation, and to identify which variables move most in constructing recourse.
  • The framework is model-agnostic: it requires only that the loss can be evaluated, not differentiability or a generative model, so it applies to black-box predictors such as tree ensembles and external APIs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The equivalence is reversible: any objective of the form loss plus regularizer can be read as a MAP problem by taking the prior proportional to the exponential of the regularizer. This means the Bayesian framing supplies a vocabulary and decision-theoretic tools, while the substantive novelty lies in the decision rules and metrics built on the posterior.
  • The posterior-mean decision is prone to landing in low-density, high-loss regions when the success region is multimodal or curved; the paper's own experiments show this failure. A natural extension is a constrained posterior mean or a decision rule that explicitly forbids low-density regions.
  • CVaR-CE depends on an assumed execution-noise distribution; matching that distribution to observed deployment noise rather than using an isotropic normal is a direct, testable refinement.
  • The ModelUnc mixture is an OR-type aggregation—it assigns mass to regions where at least one model succeeds—so it does not provide worst-case guarantees. Replacing the mixture with a loss-side aggregation such as exp(−η max_k ℓ_k) would yield a more conservative posterior, a step the paper leaves for future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a generalized-Bayes interpretation of counterfactual explanations. It defines a Gibbs posterior over counterfactuals as p(˜x|xb,y*) ∝ exp(−ηℓ(f(˜x),y*)) p(˜x|xb), chooses a distance-based prior p(˜x|xb) ∝ exp(−d(xb,˜x)), and shows in Proposition 1 that the MAP estimate equals the standard CE objective argmin[ηℓ + d]. Building on this posterior, the paper introduces two additional decision rules—a Bayes decision that minimizes posterior expected decision loss and CVaR-CE, a risk-averse rule—together with a ModelUnc construction that mixes model-specific posteriors using Bayesian weights to address model multiplicity. It defines pointwise and distributional evaluation metrics and reports experiments on 2D and 10D simulated data and on Google Trends data. The central claim is that this equivalence provides a theoretical/probabilistic foundation for conventional distance-minimization CE.

Significance. Proposition 1 is mathematically correct, and the paper is honest about several limitations, including the difficulty of MAP estimation from finite samples in high dimensions and the arbitrariness of the temperature η. The genuine contributions are the distributional decision rules (especially CVaR-CE), the model-uncertainty mixture in Eq. (8), and the distribution-level evaluation metrics. The model-agnostic nature of the framework is a practical strength. However, the paper overstates the novelty of Proposition 1: because the prior is defined as exp(−d), the equivalence is a definitional identity rather than an independent justification of distance-minimization CE. This overreach is load-bearing for the paper's stated significance and needs to be reframed. The empirical support is also limited by single-seed runs and by the acknowledged unreliability of the 10D MAP approximation.

major comments (3)
  1. [Section 2.3, Eq. (5), Proposition 1] The claimed 'theoretical basis' for distance-minimization CE is a definitional identity. The prior p(˜x|xb) ∝ exp(−d(xb,˜x)) is chosen so that the log-prior equals −d, making the MAP of the Gibbs posterior equal to argmin[ηℓ+d] by construction. Any regularized objective of the form min[ηℓ+r(xb,˜x)] can be rewritten in the same way by setting the prior to exp(−r). The paper acknowledges that the prior is user-specified, but the abstract and Section 5 nevertheless claim that this 'provides a broader theoretical foundation.' This overstates what is shown. I recommend reframing Proposition 1 as a representational equivalence—distance-minimization CE is the MAP of a Gibbs posterior under an exponential distance prior—rather than as an independent justification of the distance-minimization formulation.
  2. [Section 3.1.2, Table 3] The empirical comparison of MAP is not informative in 10D. The reported Gibbs MAP has Lpt=1.816 and Rb=0.000, which the paper attributes to the difficulty of mode estimation from finite samples using KDE/kNN. This is a candid limitation, but it means that Table 3 cannot support the claim that the approximate Gibbs MAP behaves as Proposition 1 predicts, nor can it support substantive comparisons between MAP and other decision rules in 10D. The paper should either replace the sample-based mode with a more reliable MAP estimate (e.g., direct constrained optimization of the objective or a better mode-seeking procedure) or explicitly state that the 10D MAP rows are not a validation of Proposition 1 and should not be interpreted as such.
  3. [Sections 3.1–3.2] All experiments use a single random seed (seed 42), and the thresholds and hyperparameters are hand-set (εsp/εrb = 0.25/0.35/15.0, τ = 0.9, σδ = 0.2/0.15/5.0). The paper reports means and standard deviations over base points but not over repeated seeds or over the stochastic sampling/optimization procedures. Since the paper makes quantitative claims about trade-offs (e.g., ModelUnc increases Tail and Stability, CVaR-CE improves robustness at greater distance), the absence of any repeated-seed or bootstrap uncertainty weakens the empirical conclusions. I am not demanding a full benchmark study, but a small multi-seed or bootstrap analysis would materially improve the reliability of the reported differences.
minor comments (5)
  1. [Section 3.1.2 heading] The heading reads 'Ten-Dimensional Data (Table 2)' but should refer to Table 3.
  2. [Section 3.2] The sentence 'The 2D analysis uses only x′1, summer vacation_lag1, and x′2, Jump_lag1' is awkwardly phrased; it should say the 2D analysis uses two variables, summer vacation_lag1 and Jump_lag1.
  3. [Algorithm 1] The italicized comment on line 1 ('Match the proposal distribution to the distance prior ∝ exp(−d(xb, ˜x))') is a useful clarification, but it would be better placed in the main text or as a footnote to avoid confusion about whether it is part of the algorithm's pseudocode.
  4. [Section 2.5.2] For P lu, the notation ∥˜x−xi∥ appears without a norm subscript; specifying the norm (Euclidean) at first use would improve clarity.
  5. [Appendix] The variational derivation in Eq. (20) is a standard Donsker–Varadhan result. It would be helpful to state explicitly that this derivation applies to any loss and prior and thus does not specifically justify the distance prior—this connects to Major Comment 1.

Circularity Check

1 steps flagged

The central theoretical equivalence in Proposition 1 is definitional: the distance-based prior is chosen so that the Gibbs posterior's MAP is exactly the conventional CE objective, so the claimed 'probabilistic foundation' is a re-description, not an independent derivation.

specific steps
  1. self definitional [Section 2.3, Eq. (5), Proposition 1 / Eq. (6)]
    "Choosing the distance-based prior p(˜x|xb)∝ exp(−d(xb, ˜x)) gives p(˜x|xb, y∗)∝ exp(−ηℓ(f (˜x), y∗)− d(xb, ˜x)) (5) ... Proposition 1. ... ˆxM AP = argmax ˜x p(˜x|xb, y∗) = argmin ˜x [η ℓ(f (˜x), y∗) + d(xb, ˜x)] (6) ... The distribution p(˜x|xb) is a user-specified prior over counterfactual candidates and does not represent a data-generating process."

    The posterior in Eq. (5) is constructed by placing the CE loss and the distance into an exponential; its mode is therefore, by construction, the minimizer of the same objective used in conventional CE (Eq. (1)). The prior is explicitly user-specified rather than derived, so Proposition 1 does not independently justify distance minimization from Bayesian inference. Any regularized objective min[ηL + r] can be represented as the MAP of a Gibbs posterior with prior exp(−r). The claimed theoretical foundation is thus an identity/renaming rather than a derivation that introduces new constraints or evidence.

full rationale

The main derivation in the paper is Proposition 1, and it is true only by construction: Eq. (5) defines the unnormalized posterior density as exp(−ηℓ − d), so maximizing that density is the same as minimizing ηℓ + d, which is exactly the conventional CE objective. The paper itself states that the prior is user-specified, confirming that no independent probabilistic content is added at this step. The later contributions — Bayes decision, CVaR-CE, ModelUnc, distributional metrics, and the experiments — are not part of this definitional identity and have independent content, which is why the score is not 10. However, the central claim that this equivalence provides a theoretical basis for distance-minimization CE overstates the force of the result: the equivalence is forced by the chosen prior, so the central theoretical claim reduces to a definition. Score 8 reflects that the paper's load-bearing justificatory claim is definitional, while acknowledging the presence of non-circular secondary contributions.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No fundamentally new entities; the free parameters are experimental hyperparameters. The central claim introduces no new physics or data.

free parameters (6)
  • temperature η = 1 (default)
    Controls posterior concentration; fixed to 1, sensitivity analyzed; no principled selection.
  • proposal scale σ = 1.0 (sim 2D/10D), 10.0 (real 2D), 8.0 (real 10D)
    Chosen by hand to match data scale; affects candidate spread and thus posterior samples.
  • CVaR tail probability τ = 0.9
    Chosen by hand.
  • success thresholds εsp/εrb = 0.25/0.35 (sim), 15.0 (real)
    Set by hand; success probability and robustness depend on them.
  • perturbation scale σδ = 0.2/0.15 (sim), 5.0 (real)
    Gaussian noise for robustness and CVaR; chosen without empirical calibration.
  • model weights wk = 1/K (equal)
    Equal weights used in all experiments; performance-based weights not tested.
axioms (6)
  • domain assumption Gibbs posterior p(θ|D) ∝ exp(-ηL(D,θ))p(θ) is a valid update rule (Bissiri et al. 2016).
    Section 2.2, cited; the paper relies on generalized Bayes coherence.
  • standard math exp(-d(xb,·)) is integrable for the norm-based distances used.
    Section 2.3; required for posterior to normalize.
  • domain assumption Importance sampling with a Gaussian proposal centered at xb yields a faithful approximation of the Gibbs posterior.
    Algorithm 1; for tree models the loss is piecewise constant, so the weight distribution may be degenerate.
  • domain assumption Model multiplicity is representable by a finite discrete set of models with equal weights.
    Section 2.4; reduces FG-GBCE integral to a sum.
  • standard math Donsker–Varadhan variational formula (used in appendix).
    Appendix, Eq. (19); standard.
  • standard math The decision loss ∆=||·||2 leads to posterior mean as Bayes decision.
    Section 2.5.1(b).

pith-pipeline@v1.3.0-daily-deepseek · 16547 in / 10322 out tokens · 94471 ms · 2026-08-03T14:02:19.988809+00:00 · methodology

0 comments
read the original abstract

Counterfactual explanations (CEs) enhance the interpretability of machine learning models by identifying the smallest change to an input required to obtain a desired output. Although CEs are conventionally formulated as a distance-minimization problem, the theoretical basis of this formulation has received limited attention. We show that a distance-minimization-based CE is mathematically equivalent to the maximum a posteriori (MAP) estimate of a Gibbs posterior within the generalized Bayes framework, specifically when a distance-based prior is used. We call this formulation the Distance-Prior Generalized Bayes CE (DP-GBCE). Building on this posterior perspective, we introduce two decision rules beyond MAP within a unified framework: a Bayes decision that minimizes expected decision loss and CVaR-CE, a risk-averse decision rule. We also propose an extension that uses Bayesian model weights to mix the posterior distributions of multiple models, thereby accounting for model multiplicity, where several models have comparable predictive performance. Finally, we define metrics for evaluating both individual CEs and the posterior distribution as a whole, and use experiments on simulated data and Google Trends data to quantify the trade-offs among the decision rules.

Figures

Figures reproduced from arXiv: 2607.29077 by Keita Kinjo.

Figure 1
Figure 1. Figure 1: 2D Gibbs posterior and CE decisions (MAP, Mean, CVaR-CE, DirectOpt [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: 2D model-uncertainty CE posterior and CE decisions (MAP, Mean, CVaR-CE, [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: η-sensitivity analysis: mean ± SD of stability (left) and success probability (right) (Gibbs posterior; 5 randomly chosen base points xb) [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: η-sensitivity analysis: mean CE loss (MAP, Mean, CVaR-CE) [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 2 linked inside Pith

  1. [1]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    A general framework for updating belief distributions , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2016 , publisher=

    2016

  2. [2]

    arXiv preprint arXiv:1901.05353 , year=

    A primer on PAC-Bayesian learning , author=. arXiv preprint arXiv:1901.05353 , year=

    Pith/arXiv arXiv 1901

  3. [3]

    Data Mining and Knowledge Discovery , volume=

    Counterfactual explanations and how to find them: literature review and benchmarking , author=. Data Mining and Knowledge Discovery , volume=. 2024 , publisher=

    2024

  4. [4]

    ACM Computing Surveys , volume=

    Counterfactual explanations and algorithmic recourses for machine learning: A review , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    2024

  5. [5]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , pages=

    Robust counterfactual explanations in machine learning: a survey , author=. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , pages=

  6. [6]

    arXiv preprint arXiv:2501.05795 , year=

    Robust counterfactual explanations under model multiplicity using multi-objective optimization , author=. arXiv preprint arXiv:2501.05795 , year=

    Pith/arXiv arXiv

  7. [7]

    Uncertainty in Artificial Intelligence , pages=

    Robust bayesian recourse , author=. Uncertainty in Artificial Intelligence , pages=. 2022 , organization=

    2022

  8. [8]

    Uncertainty in Artificial Intelligence , pages=

    On counterfactual explanations under predictive multiplicity , author=. Uncertainty in Artificial Intelligence , pages=. 2020 , organization=

    2020

  9. [9]

    International Conference on Artificial Intelligence and Statistics , pages=

    Bayesian hierarchical models for counterfactual estimation , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2023 , organization=

    2023

  10. [10]

    International Conference on Artificial Intelligence and Statistics , pages=

    Exploring counterfactual explanations through the lens of adversarial examples: A theoretical and empirical analysis , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2022 , organization=

    2022

  11. [11]

    2013 , publisher=

    Counterfactuals , author=. 2013 , publisher=

    2013

  12. [12]

    Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency , pages=

    Explaining machine learning classifiers through diverse counterfactual explanations , author=. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency , pages=

    2020

  13. [13]

    Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency , pages=

    Actionable recourse in linear classification , author=. Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency , pages=

    2019

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Algorithmic recourse under imperfect causal knowledge: a probabilistic approach , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    ACM Computing Surveys , volume=

    A survey of algorithmic recourse: contrastive explanations and consequential recommendations , author=. ACM Computing Surveys , volume=. 2022 , publisher=

    2022

  16. [16]

    Counterfactual explanations without opening the black box: Automated decisions and the

    Wachter, Sandra and Mittelstadt, Brent and Russell, Chris , journal=. Counterfactual explanations without opening the black box: Automated decisions and the