acceptodds
Under review as a conference paper at ICLR 2027

Noise-Robust Counterfactual Explanations under the Rashomon Effect via Homotopy Path-Following

Abstract

Although there has been extensive research on robust counterfactual explanations (RCEs) addressing the Rashomon Effect, when training data contains noisy labels, the Hessian matrix — which characterises the geometry of the ellipsoid — becomes corrupted under label noise for RCE methods based on Rashomon uncertainty sets. This corruption distorts the geometry of the uncertainty set, causing robust optimisation to yield unreliable RCEs. We find that the perturbation introduced by noisy labels to the Hessian matrix exhibits a low-rank structure, leading to a deflection of the principal directions of the Rashomon ellipsoid defined by the Hessian; this phenomenon is observable across different noise rates, regardless of whether the label noise is random flipping or systematic flipping. Our theoretical analysis further establishes that this conclusion generalises to model uncertainty sets defined by general quadratic forms. To mitigate the impact of noisy labels on RCE, we propose a path-following strategy based on homotopy: starting from a conservative uncertainty set that is independent of labels, label information is gradually incorporated along a continuously deformed path, with the homotopy parameter controlling the degree of trust placed in the training labels. We prove that the robust CE solution is continuously differentiable along the homotopy path, enabling stable path-following and principled selection of the homotopy parameter. Experimental results on five diverse datasets demonstrate that this new framework produces counterfactual explanations that are both robust and computationally efficient across model uncertainty sets defined by different quadratic forms.

Primary area:
interpretability and explainable AI
Keywords:
Counterfactual Explanations, Label Noise, Rashomon Sets

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UPwYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies robust counterfactual explanations under the Rashomon effect when training labels are noisy. Building on ElliCE, which approximates the Rashomon set using a Hessian-defined ellipsoid, the authors show empirically that label noise can distort this uncertainty set by inducing a structured, low-rank perturbation that rotates its principal directions. To mitigate this, they propose a Homotopy Path-Following (HPF) method that interpolates between a label-agnostic conservative uncertainty set and the label-conditioned one using a scalar homotopy parameter. They further extend the construction to general quadratic uncertainty sets and show continuity of the robust counterfactual solution along the homotopy path. Experiments on five datasets, under symmetric and asymmetric synthetic label noise, compare HPF with ElliCE and other robust counterfactual baselines, and report improved robustness with competitive recourse cost and computational efficiency.

Critical strengths
  • The paper identifies an important failure mode of Rashomon-set-based robust counterfactual methods under label noise, where the uncertainty set itself can become distorted.

  • HPF provides a simple correction with theoretical continuity guarantees and is evaluated across five datasets and multiple baselines, showing strong robustness and competitive efficiency.

Critical weaknesses
  • The method is closely built on ElliCE, with the main addition being a scalar homotopy between a label-agnostic and label-conditioned uncertainty set. I find this extension somewhat incremental relative to the ICLR bar.

  • The proposed correction is motivated and validated for label-noise-induced geometric distortion. It remains unclear how much of the method’s benefit extends to other forms of uncertainty-set misspecification, which limits the broader robustness claims.

Questions for authors
  1. How dependent is HPF on the specific low-rank geometric distortion induced by label noise? Would the method still provide robustness if the uncertainty set were misspecified by a qualitatively different shift, e.g., one that does not primarily rotate a few leading directions?

  2. Have the authors evaluated the sensitivity of HPF to the homotopy parameter tt, e.g., by comparing performance across fixed values of tt? Since t controls the trade-off between the label-agnostic and label-conditioned uncertainty sets, such an analysis would help clarify how much of the performance gain comes from the learned choice of t and how sensitive the method is to this choice.

Reviewer_Q7FDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies robust counterfactual explanations (CEs), meaning CEs that stay valid for every model in a set of near-optimal models. It builds on ElliCE (Turbal et al. 2026), which approximates this set by an ellipsoid around the trained last-layer weights, shaped by the Hessian of the training loss. The authors argue that noisy training labels distort this ellipsoid. In an empirical study on five tabular datasets, they report that label noise keeps the ellipsoid low-rank (effective rank about 2 for the main network) but rotates its main axis by about 4 to 19 degrees, and that the change in the Hessian is dominated by a single direction. To reduce this effect, they propose HPF, which replaces the Hessian with a weighted average of a label-independent matrix (obtained by setting all predicted probabilities to 0.5) and the usual Hessian, with weight t on the Hessian. The weight t is chosen by gradient ascent on a robustness objective, with a lower bound of 0.6. The authors prove that the CE changes smoothly with t (Theorem 2) and extend the approach to matrices based on the Fisher information and on bootstrap estimates. HPF is compared with four baselines (ElliCE, ROAR, T-Rex, PROPLACE) on five datasets with 10% injected label noise, using validity, robustness to model changes (the share of CEs that stay valid under 100 alternative models), L2 distance, and LOF. Across both robustness evaluators and all three label settings (clean, symmetric noise, asymmetric noise), the best HPF variant has the highest or joint-highest robustness on all five datasets. However, ElliCE already reaches a robustness of at least 0.91 on four of the five datasets, so the clear robustness gain is on COMPAS only; on the other four, HPF matches ElliCE's robustness with shorter or similar distances, while ROAR and T-Rex give the shortest distances overall.

Critical strengths
  • The problem is timely and important: robust CE methods assume the estimated set of near-optimal models is reliable; noisy labels are common in practice and haven't been studied before in the robust CE context.
  • The spectral study, showing that label noise rotates the Hessian ellipsoid along a few directions, is an informative observation and is repeated across five datasets, two noise types, several noise rates, and three network depths.
Critical weaknesses
  • The evaluation does not seem to test the paper's central claim. The paper's setting assumes the training labels are noisy, but its argument requires robustness relative to the uncertainty set obtained from clean labels: the spectral study itself measures the rotation of the noisy Hessian against the clean one, and if the target were the noisy-label set, ElliCE's ellipsoid would already be the right approximation and there would be nothing to correct. Since the noise is injected synthetically, clean labels are available for this test. Instead, both robustness evaluators build their alternative models from the same noisy training data (retraining on the same training fold, or perturbing the trained model within a loss tolerance computed on the noisy labels), so they measure robustness to the very noisy-label uncertainty the method claims is misleading. Consistent with this, HPF's advantage over ElliCE is about the same with clean labels as with noisy ones, and only one noise rate (10%) is tested. This is the main reason for the rating: the experiments do not show that HPF corrects the effect of label noise.
  • The method and theory do not seem to match the claim of greater robustness. For the Hessian variant, the label-independent matrix is always at least as large as the Hessian (since σ(1−σ) ≤ 1/4), so every set along the path lies inside ElliCE's set: HPF asks for less robustness than ElliCE, not more. The robustness objective in Eq. 6 is always highest for the smallest allowed set, so t lands at the lower bound of 0.6, and the shorter distances on four datasets are what this weaker requirement would produce. (The Fisher and bootstrap variants are rescaled to the same size as the label-independent matrix, so their sets differ in shape rather than strength; for them the robustness claim is untested rather than contradicted, as noted above.) On the theory side, the proof of Theorem 2 assumes the constraint is concave in the input, which does not hold for the MLPs studied, and the generalization to quadratic uncertainty sets claimed in the abstract is not proven. This is the second reason for the rating: the paper's explanation for why HPF should be more robust is not supported by its method or its proofs.
Questions for authors
  • Do the retrained and perturbed models in your evaluators use the noisy labels? If so, what uncertainty set is HPF meant to be robust to, given that ElliCE's ellipsoid already approximates the set for the noisy data? Since the noise is injected synthetically, can you measure robustness against models trained on the clean labels (or at least on an independent draw of noisy labels), and show that HPF's advantage over ElliCE grows with the noise rate?
  • In the spectral study, how are the clean and noisy Hessians obtained (separately trained networks, same seed?), and how much does the main axis rotate between two clean models trained with different seeds? Since HPF interpolates globally, how does it use the low-rank structure found in this study?
Reviewer_CfvFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces HPF (Homotopy Path-Following), a method for generating robust counterfactual explanations when training labels are noisy. Based on a Hessian spectral analysis, the authors argue that label noise rotates the Rashomon ellipsoid used by methods such as ElliCE. To reduce this effect, HPF interpolates between a label-agnostic matrix and the label-dependent Hessian, with a single parameter t controlling how much the labels are trusted. The authors also claim that the solution varies smoothly along this path, extend the approach to general quadratic uncertainty sets, and report high validity and robustness at competitive cost on five tabular datasets.

Critical strengths
  • Paper considers a new type of problem (noisy labels and their influence) that has not discussed in robust counterfactual explanations literature so far
  • The proposed method is fast and valid
Critical weaknesses

W1. The motivation for the work seems a little artificial to me: how real is the problem caused by the noise label? How much does it influence the robustness of the existing methods? A more thorough analysis of this influence is required before addressing the issue. The obtained results undermine the need for a specialized method targeting this issue: on four of the five datasets, the robust fraction of ElliCE drops by at most 0.03 under AWP and at most 0.07 under Retrain when noise is injected (Tables 7 and 8), and its low score on COMPAS is already present on clean labels (0.72 and 0.60).

W2. No code is shared, which is important when there are doubts regarding the quality of the experiments, see W4.

W3. The proposed method builds strongly on the ElliCE method by Turbal et al.: it changes only the matrix that defines the ellipsoid, and setting t=1t = 1 recovers ElliCE itself. The method also inherits the ElliCE weakness – it’s restricted to just MLPs and linear models, which, compared to decision-tree-based methods, are a less common choice for tabular data tasks. In addition, no counterfactual desiderata other than robustness (validity indirectly) and proximity are considered in the formulation of the method; plausibility is only measured afterwards.

W4. The reference methods seem to be inadequately evaluated. T-Rex reaches validity of only 0.32 on FICO and 0.59 on German, and ProPlace only 0.43 on COMPAS and 0.09 on FICO, while their original papers report over 96% on HELOC (FICO) and German (Hamman et al., 2023) and 100% on COMPAS and HELOC (Jiang et al., 2024), even on retrained models. ElliCE’s validity of 0.72 on COMPAS is also surprising, because its constraint implies validity. ROAR on German with symmetric noise has validity 0.97 and L2 3.76 in Table 7, against 0.29 and 1.42 in Table 8, although validity and L2 are computed on the original model and should be the same in both tables.. The baseline hyperparameters are not reported, so these results cannot be checked.

W5. The paper has numerous grammar, spelling, spacing and factual errors. Theorem 2 is provided while Theorem 1 never appears anywhere, and the statement of Theorem 2 has an empty equation reference. There are inconsistencies between the text and the results reported in tables: for example, the effective rank on FICO is 1.98–2.25 in Section 3 but 2.09–2.21 in Table 3. Some results are also duplicated across datasets. In Table 7, the asymmetric-noise results of HPF (FIM) on FICO are identical to those on German (L2 6.70 ±\pm 1.58, LOF 1.01 ±\pm 0.07), and the same FICO values appear in Table 1, while Table 8 reports L2 3.92 for that setting. The asymmetric-noise results of HPF (Bootstrap) on Diabetes are identical to those on Australian (L2 2.45 ±\pm 0.81, LOF 1.12 ±\pm 0.09). In Section 5.1 there is a sentence: “ROAR performs best under clean labels [...] but its reliance on nearest-neighbour structures” – ROAR does not use neighbour structures at all. Substantial proofreading is required, and the reported results should be rechecked.

Questions for authors
  1. How are the reference methods evaluated? Can you provide details on that, share code?
  2. What differentiates substantially your method from ElliCE?
  3. Can you further motivate the need for the proposed method?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.