- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes a distributionally robust end-to-end learning framework, which predicts the probability distribution of uncertain parameters through a conditional generative model and constructs a Wasserstein ambiguity set using Sinkhorn divergence and conformal calibration. Theoretically, by correcting entropy regularization and finite-sample estimation errors, the authors establish a finite-sample marginal coverage guarantee that the ambiguity set contains the true conditional distribution, and utilize KKT implicit differentiation to achieve end-to-end optimization of distribution prediction, ambiguity set calibration, and downstream DRO. Finally, several experiments demonstrate that this method reduces decision regret compared to the selected baseline.
From a methodological design perspective, I believe this paper offers very limited methodological innovation. The methodological modules involved, such as end-to-end learning of fuzzy set parameters, using generative models to provide contextual distribution samples, and using conformal prediction to handle the context shift problem, are all readily available in similar papers. The overall workflow framework is very similar to "Robust Decisions via Generative Wasserstein Distributionally Robust Optimization," with the core difference being the use of the same radius for all contexts and the use of Sinkhorn divergences as divergences as nonconformity scores to provide marginal coverage results. This is the paper's core theoretical innovation.
However, the problem lies in the fact that marginal coverage cannot provide the probability guarantee for the specific context that contextual DRO is most concerned with. The average probability guarantee cannot handle extreme context situations. Furthermore, the introduction of Sinkhorn divergences brings additional conservatism compared to Wasserstein distance. Therefore, compared to contextual SO, the current marginal results do not provide a more meaningful statistical guarantee, and the performance improvement compared to contextual SO is merely empirical. I look forward to the authors providing further theoretical results comparing the performance improvement of contextual SO.
Finally, the analysis of the preconditions for KKT implicit differentiation in the paper is insufficient. The convexity and Slater conditions assumed in the main text alone cannot guarantee that the optimal solution maps to differentiable paths. Furthermore, in the actual derivation in Appendix J, the authors use stronger conditions such as nonsingular KKT matrices, which should be reflected in the main text.
Critical strengths
I believe this paper has only one critical strength: using Sinkhorn divergences as nonconformity scores. This solves the difficulty of differentiating the Wasserstein distance, but also introduces additional conservatism. The marginal coverage result obtained through this innovation is the core theoretical difference compared to similar pipeline papers.
Critical weaknesses
As I mentioned in the paper assessment, the paper's main weakness lies in its limited methodological novelty compared to similar e2e DRO papers, and its conservative approach of using the same radius for ambiguity sets under all contexts. The presented marginal theoretical results do not address the conditional coverage for individual contexts, which is particularly important in contextual decision-making. Furthermore, it does not establish any theoretical advantage in decision performance or regret over contextual stochastic optimization. The reported performance improvements over the contextual SO baselines are supported only by empirical results.
Questions for authors
Could the authors provide a theoretical analysis demonstrating the performance advantages of the proposed method over end-to-end contextual stochastic optimization (SO)? I am concerned that using a context-independent radius for all Wasserstein ambiguity sets may introduce excessive conservatism, potentially undermining the performance gains achieved through end-to-end learning. Can the authors theoretically characterize this trade-off and establish conditions under which the proposed method outperforms contextual SO in terms of decision regret?
4 comments
Response to Reviewer tjF3 (Part 2)
Theoretical comparison with contextual SO
If the learned law is exact and there is no distribution shift, SO is optimal and no method can improve over it. We therefore characterize the trade-off in two regimes and give the condition under which DRDFL outperforms SO.
Without shift (new Remark J.1). Let be the model error, which our calibration bounds by with probability at least . For , the DRDFL decision minimizes the SO objective plus , where is the sensitivity of the loss to (Eq. (10), Proposition J.1). The true and predicted expected losses of any decision differ by at most . On the coverage event, this gives:
- Guarantee. The true expected loss of is at most the DRO value (Theorem 4.1), and its regret is at most , where is the true optimal decision, because the DRO value is at most the robust objective at . For the SO decision , the worst-case loss over the calibrated ambiguity set is its predicted loss plus , and is the decision that minimizes this worst case.
- Cost. If the learned law is exact, the excess loss of DRDFL over SO is at most .
DRDFL is therefore the better choice when the model error is adverse to the SO decision, at a cost of at most the radius times the reduction in sensitivity when the model is exact.
With shift (new Appendix P). DRDFL (ours) performs better than end-to-end contextual SO when there is a distribution shift between the training data and the test data, especially with a large distribution shift. We place SO in its most favorable position: it knows the true conditional distribution of the training data exactly, while DRDFL only has its prediction. Theorem P.1 gives an exact threshold on the shift between the training and the test distribution. Whenever the DRDFL decision is less sensitive than the SO decision (), its worst-case expected loss over all shifts of size at most is smaller than that of SO if and only if , the loss given up without shift divided by the reduction in sensitivity. When the ambiguity set covers the model error, , and if the predictor is exact. The comparison is stated for the worst case because a particular shift may favor either decision.
Context-independent radius. The threshold holds feature by feature, so a common radius does not weaken it. On average over the features, the same threshold holds with averaged quantities, and the features that fails to cover only add a term proportional to the miscoverage, which is at most on average (Theorem P.1(iii)).
The experiments confirm. DRDFL is better under shift and SDFL is better without shift (Section 5.3).
Marginal Coverage vs. Conditional Coverage
We agree with the reviewer on that a radius with a coverage conditioned on each feature can results in a better performance than a radius with a marginal coverage (averaged over all features).
Post-training conditional coverage. We propose a decision-focused calibration algorithm which can be used after the end-to-end training finishes and gives a coverage guarantee conditioned on each feature (Theorem N.1). See Appendix N Decision-Focused Calibration with a Conditional Coverage Guarantee for details.
Impossibility of conditional coverage in end-to-end training. However, this algorithm is not diferentiable and cannot be integrated in the end-to-end training. We even show that under the current setting having conformal calibration with a conditional coverage in the end-to-end training is impossible. See for Appendix N.1 for details.
Conditions for KKT differentiation
Section 4 now states the conditions under which we differentiate the KKT system (Assumption K.1): linearly independent active constraint gradients, strict complementarity, positive definiteness of on their null space, and smoothness of . Under these conditions, the solution is locally unique and differentiable [2].
Response to Reviewer tjF3 (Part 3)
Novelty of Our Paper and Framework
Aware of the true uncertainty. Compared to frameworks like Gen-WDRO, whose learnable radius cannot quantify the uncertainty properly, our framework (DRDFL) proposes a true uncertainty-aware end-to-end training of a distribution predictor.
Robust to distribution shift. Compared to stochastic decision-focused learning and Gen-WDRO, whose performance degrades under a distribution shift, ours learns this uncertainty via conformal calibration and performs robustly.
Bounds and coverage. We derive explicit one-sided finite-sample bounds that relate empirical debiased Sinkhorn scores to population and discrepancies. Using these bounds, we calibrate Wasserstein ambiguity sets with finite-sample marginal coverage of the true conditional law in an end-to-end decision-focused learning framework.
References
[1] Foygel Barber, Rina, et al. "The limits of distribution-free conditional predictive inference." Information and Inference: A Journal of the IMA 10.2 (2021): 455-482.
[2] Fiacco, Anthony V. "Sensitivity analysis for nonlinear programming using penalty methods." Mathematical programming 10.1 (1976): 287-311.
[3] Xu, Han, and Christopher Yeh. "Robust Decisions via Generative Wasserstein Distributionally Robust Optimization." NeurIPS 2025 Workshop: Generative AI in Finance.
We would be happy to discuss, expand, or provide any further clarifications that the reviewer may consider beneficial.
A Gentle Reminder
Dear reviewer,
Did our rebuttal address your concerns? Please let us know. We are happy to add any further clarifications.
The end of the discussion period is quickly approaching.
Thank you.
Response to Reviewer tjF3 (Part 1)
We thank the reviewer for the time, the thoughtful comments, and the helpful suggestions for improving our paper. We directly address the questions and the raised weaknesses.
Comparisons with Gen-WDRO and the impact of conformal radius
To our best knowledge, Gen-WDRO is the only published paper studying a similar problem and using task-learned radius selection. We implemented the framework of Gen-WDRO according to its paper. We compare our DRDFL against it in both synthetic and real-world experiments, which can be found in Table 1 (DRDFL outperforms Gen-WDRO).
Why Gen-WDRO fails?
In experiments, the learned radius of Gen-WDRO decreases to a very small value at the end of the end-to-end training, which comes from using the same data to train the CNF and the neural network for predicting the radius in parallel shown in Figure 1 of [3]. This makes Gen-WDRO's performance very close to stochastic DFL. When there is an obvious distribution shift between the training data and the test data, Gen-WDRO and stochastic DFL are unable to capture such uncertainty and shift and perform worse than DRDFL.
In Gen-WDRO the radius is merely trained to minimize the task loss rather than to contain the true distribution. For the problem studied in [3], there can be multiple combinations of the learned distribution's mean and the learned radius giving the same objective value, which is also mentioned in [3] by the authors. The ambiguity set with the learned radius in Gen-WDRO cannot be viewed as the true uncertainty.
Why conformal radius? Any other framework training a distribution predictor and a radius predictor in parallel using the same data and the task loss, the 2 problems of Gen-WDRO mentioned above will occur. To resolve these 2 problems, different data from the one for learning the distribution has to be used to quantify the uncertainty of the predicted distribution in each training iteration. Conformal calibration is the simplistic realization of this and can be differentiated. See Appendix O Comparison with Gen-WDRO for more details.
Also, a fixed radius can simply result in very conservative performance. If a user sets a too large fixed radius, whose effect is similar to a too high confidence level, the regularization coming from the ambiguity set will decrease the performance very much as shown in Figure 4 and 5. In contrast, a too small fixed radius cannot capture the uncertainty and the distribution shift.
Marginal coverage and the common radius. Distribution-free coverage conditional on each context is not attainable in finite samples [1], and contextual SO and Gen-WDRO provide no coverage guarantee of any kind. For a given context, the newly added Appendix N gives a guarantee on the decision that is conditional on that context (Theorem N.1); it applies when outcome samples at the context are available. The common radius is a quantile of the predictor's own errors, so it shrinks as the predictor improves, and the confidence level controls its conservatism (Figures 4–5). With this radius, DRDFL has 17% and 47% lower regret than end-to-end contextual SO (SDFL) with the same predictor (Table 1).
Conservatism of the Sinkhorn score.
In Table 2 (Appendix M.1), we compare the Wasserstein distances between 2 distribution, the Sinkhorn divergences with and without correction terms. It shows that the Sinkhorn divergences with 2 correction terms does not exceed the true Wasserstein distance by a large margin with a proper entropic regularization weight. For high-dimensional distributions, the majority of the "overestimation" comes from the plug-in correction not the entropic correction. For real-world experiments, where only finite samples from distributions (no closed forms of the distributions) are available, the best we can target on is the Wasserstein distances using the plug-in estimators (empirical measures averaged over finite samples). Therefore, the plug-in correction is not added and the "overestimation/conservatism" is very little. See Section 5.1 and Appendix M.1 for more details.