acceptodds
Under review as a conference paper at ICLR 2027

Contextual Distributionally Robust Optimization with Causal and Continuous Structure

Abstract

We propose a framework for contextual distributionally robust optimization (DRO) that considers the causal and continuous structure of the underlying distribution, and we develop interpretable and tractable decision rules. We first introduce the causal Sinkhorn discrepancy (CSD), an entropy-regularized causal Wasserstein distance that encourages continuous transport plans while preserving causal consistency. We then formulate a contextual DRO model with a CSD-based ambiguity set, termed Causal Sinkhorn DRO (Causal-SDRO), and derive its strong dual reformulation, where the worst-case distribution is characterized as a mixture of Gibbs distributions. To obtain an (infinite-dimensional) optimal policy, we propose a soft regression forest (SRF) decision rule: it preserves the interpretability of classical decision trees while being fully parametric, differentiable, and Lipschitz-smooth, enabling intrinsic interpretation from both global and local perspectives. To solve the Causal-SDRO with parametric decision rules, we develop an efficient stochastic compositional gradient algorithm. Finally, we validate our method through numerical experiments on synthetic and real-world datasets, demonstrating its superior performance and interpretability.

Primary area:
optimization
Keywords:
contextual distributionally robust optimization, causal Sinkhorn discrepancy, soft regression forest, stochastic compositional optimization, interpretable decision-making

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Rt4hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper combines causal transport constraints with entropy regularization to learn robust decisions from contextual data. The authors derive a dual formulation and a Gibbs-mixture expression for the worst-case distribution. They use a soft regression forest (SRF) to represent the decision rule and adapt a stochastic compositional gradient method to train it, with a claimed O(ϵ−4)O(\epsilon^{-4}) complexity for reaching an ϵ\epsilon-stationary point. The method is evaluated on newsvendor, inventory-substitution, and portfolio problems, with comparisons across DRO models and policy classes.

Critical strengths
  1. A useful connection between causal transport and entropy regularization. I appreciate that Eqs.~ (3a) -- (3b) make the role of the conditional outcome distribution explicit. When several outcomes share the same source covariate, requiring them to use the same covariate transport rule is a meaningful restriction. The worst-case density in Section~3.2 also helps explain which scenarios the model protects against. I see value in this formulation for problems where that conditional variation is represented in the source distribution.

  2. A practical policy representation and relevant applications. I find the SRF a reasonable choice for learning decisions directly from covariates. Its output remains feasible for a convex decision set if the leaf decisions are feasible, and its differentiability allows the splits and leaf decisions to be trained together. I also appreciate the comparisons of both DRO models and policy classes, including a portfolio study using real data.

Critical weaknesses
  1. The causal constraint is inactive when covariates are unique. My main concern is that Section 5.1 groups observations by exact covariate matches. If each covariate vector appears once, then Y^\hat Y is determined by X^\hat X under P^\hat P, so every coupling satisfies X⊥Y^∣X^X\perp\hat Y\mid\hat X. Hence Γc(P^,P)=Γ(P^,P).\Gamma_c(\hat P,P)=\Gamma(\hat P,P). Causal-SDRO then coincides with SDRO under matched costs, reference measures, and parameters. Appendix F.1 describes Gaussian covariates but does not specify whether contexts are repeated. With an independent draw per observation, covariates are unique almost surely. I would appreciate clarification of any grouping or conditional estimation used, and an explanation of the performance differences in Figure 3. Without this, I cannot attribute the reported gains to the causal constraint.

  2. The convergence proof does not clearly cover Algorithm 1. I cannot reconcile Algorithm 1 with the conditional structure of Eq. (8). Its gradient requires normalization separately for each x^\hat x, and pooling the conditional expectations can change the gradient. Yet Algorithm 1 resamples x^\hat x and ξ1\xi_1 while maintaining one shared scalar y1ky_1^k and vector y2ky_2^k. The proof in Appendix G.4 instead maintains functions y1k(x^)y_1^k(\hat x) and y2k(x^,ξ1)y_2^k(\hat x,\xi_1) for every condition (Eqs. (45)–(46)). Since ξ1\xi_1 has continuous support, I would appreciate an explanation of how these functions are represented and evaluated, and how their cost is included in Theorem 6. Until this gap is resolved, I am not convinced that the stated algorithm has the claimed convergence and complexity guarantees.

Questions for authors

1.Are covariates repeated or grouped in the newsvendor experiments? If every n(x^)=1n(\hat x)=1, could the authors check whether Causal-SDRO and SDRO give the same objective and gradient at identical parameters, and explain the differences in Figure~3?

  1. How does the implementation estimate the expectations separately for each (x^,ξ1)(\hat x,\xi_1)? Could the authors reconcile Algorithm1 with Eqs.(45)--(46), including how the conditional estimates are represented and how their evaluation cost enters Theorem~6?
3 comments

We very much appreciate the reviewer’s invaluable comments on our manuscript and now provide our point-by-point response. The revised parts of our manuscript are highlighted in blue\textcolor{blue}{\text{blue}}.

  • [How are empirical conditional distributions constructed?] In revision, we clarify the number of uncertain parameters associated with each covariate, i.e., nm∈1,5,10n_m \in \\{1, 5, 10\\}, in synthetic experiments (newsvendor and inventory substitution problems). For the portfolio problem with real-world data, we employ a k-nearest neighbors (k-NN) approach to group similar covariates, each group containing nm∈1,10n_m\in\\{1,10\\} observations.

It is true that when only one uncertain parameter is associated with one covariate, Causal-SDRO and SDRO are equivalent. In our previous experiment results, the observed performance discrepancy between Causal-SDRO and SDRO was due to differences in the computational algorithms and hyperparameter choices. In the revised experiments, we use identical hyperparameters for Causal-SDRO and SDRO. Moreover, since the two formulations are equivalent when nm=1n_m=1, we only solve Causal-SDRO and use the resulting solution for both methods, ensuring identical numerical results in this case. See the Table in our general response to the whole reviewer team for the example of newsvendor, and detailed box plots for other experiments are provided in the manuscript (Figures 3, 9, 11)

  • [How does the proposed algorithm work? Are the existing convergence results for the algorithm inconsistent with the algorithm itself? ] In the revision (Section 5.2), we developed a new algorithm for our problem, termed CM-SCSC, that maintains vector states to track the nested expectation terms. Specifically, at each iteration, we draw QQ i.i.d. samples of ξ2\boldsymbol{\xi}_2 from WϵW_\epsilon, to estimate the inner expectation with respect to ξ2\boldsymbol{\xi}_2 conditional on each historical covariate value x^m\widehat{\boldsymbol{x}}_m and sampled ξ1\boldsymbol{\xi}_1. Each estimate is an NN-dimensional vector, where NN is the number of all historical observations. The outer expectation of ξ1\boldsymbol{\xi}_1 conditional on x^m\widehat{\boldsymbol{x}}_m is tracked by a scalar variable for each covariate group. These scalar variables are stored as an MM-dimensional vector, where MM is the number of distinct observed covariate values. We have also corrected the theoretical analysis accordingly (See Theorem 6 in Section 5.2, with the proof in Appendix H.5).

Thank you for the detailed response and revisions. The corrected experiments clarify the equivalence with SDRO when each context has one observation and show gains from the causal constraint when contexts have multiple observations. The revised algorithm and proof address my main concern about how the conditional expectations are estimated, although the computational cost as the dataset grows could be made clearer. In light of these revisions, I am raising my overall rating to 3.

Thank you for your feedback. We want to make an additional remark regarding the computational cost as the dataset grows.

  • In Theorem 6 and Remark 7 of our revised manuscript, we characterize the number of iterations and per-iteration computational cost (quantified as the number of scalar function evaluations) in terms of MM (number of distinct observed covariate values) and NN (total number of historical observations). The number of iterations is O(ε−4)\mathcal{O}(\varepsilon^{-4}) and independent of M,NM, N. The per-iteration computational cost is O(MNε−1)\mathcal{O}(MN\varepsilon^{-1}).

  • We provide additional experimental studies on the newsvendor problem to examine the influence of MM and NN on the actual training time. In Table 1, we fix N=4000N=4000 and increase MM by decreasing the number of observations associated with each covariate. In Table 2, we fix M=40M=40 and increase NN by increasing the number of uncertain observations associated with each covariate. In Table 1, we find the training time increases with MM for fixed NN. However, the results in Table 2 indicate the training time remains relatively stable for N≤4000N\le 4000 for fixed MM. The reason is that our implementation leverages GPU that performs parallel computing to implement Steps 5 and 7 of Algorithm 1, so that the actual training time has smaller dependence on NN.

Table 1. Training time (s) with fixed N=4000N=4000 and increasing MM (averaged over 3 trials).

MM408020040080020004000
1-Causal-SDRO Time (s)7.7288.0278.0407.9697.8648.79111.582
2-Causal-SDRO Time (s)7.1317.5167.1397.2427.8068.13710.956

Table 2. Training time (s) with fixed M=40M=40 and increasing NN (averaged over 3 trials).

NN404008001600240032004000
1-Causal-SDRO Time (s)7.6547.7467.7467.9137.8067.9117.728
2-Causal-SDRO Time (s)6.7497.1146.8967.1597.1767.1717.131
Reviewer_qSZuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Causal-SDRO, a contextual distributionally robust optimization framework combining causal transport constraints with entropy regularization. It derives a nested dual formulation and Gibbs-type worst-case distributions, introduces a soft regression forest (SRF) policy, and proposes a stochastic compositional optimization algorithm. Experiments cover newsvendor, inventory substitution, and portfolio allocation problems.

Critical strengths
  • Combining causal transport with entropy regularization is a coherent modeling direction. When the reference distribution has nondegenerate conditional outcomes, the nested formulation captures a meaningful restriction on adversarial feature transport.
  • The paper establishes strong duality results for the resulting DRO problem. The proof seems to be correct after a brief read.
  • The paper considers three decision problems, compares several DRO formulations using the same SRF policy class, and examines policy architecture and hyperparameter sensitivity.
Critical weaknesses
  • The paper should more clearly distinguish its contributions from methods it already cites. SCSC is an existing optimization algorithm, and the proposed EIG attribution is expected gradients using training examples as reference points. Comparisons with simple noise-augmentation or smoothing baselines would also help show whether the gains come specifically from the proposed DRO formulation.
  • The proposed causal constraint has no effect when every training example has a different feature vector, as expected under the stated synthetic data generation. In that setting, Causal-SDRO should reduce to ordinary SDRO, making the reported performance differences difficult to explain.
Questions for authors
  • How many distinct feature vectors occur in each dataset? If contexts are unique, how do you explain the reported Causal-SDRO/SDRO differences?
  • Which experimental losses satisfy Assumption 2? If smoothing or bounded approximations are used, please provide their definitions and approximation errors.
  • How is the p=1 portfolio objective kept finite? With unbounded returns, do you restrict the return support or truncate perturbations?
  • How do the experiments satisfy the convergence assumptions?
3 comments

We are deeply grateful for your comments. We now provide point-by-point responses, and the revised parts of our manuscript are highlighted in blue\textcolor{blue}{\text{blue}}.

  • [How many distinct feature vectors occur in each dataset?] We clarify the number of uncertain parameters associated with each covariate, i.e., nm∈1,5,10n_m \in \\{1, 5, 10\\}, in synthetic experiments (newsvendor and inventory substitution problems). For the portfolio problem with real-world data, we employ a k-nearest neighbors (k-NN) approach to group similar covariates, each group containing nm∈1,10n_m\in\\{1, 10\\} observations.

It is true that when only one uncertain parameter is associated with one covariate, Causal-SDRO and SDRO are equivalent. In our previous experiment results, the observed performance discrepancy between Causal-SDRO and SDRO was due to differences in the computational algorithms and hyperparameter choices. In the revised experiments, we use identical hyperparameters for Causal-SDRO and SDRO. Moreover, since the two formulations are equivalent when nm=1n_m=1, we only solve Causal-SDRO and use the resulting solution for both methods, ensuring identical numerical results in this case. See the Table in our general response to the whole reviewer team for the example of newsvendor, and detailed box plots for other experiments are provided in the manuscript (Figures 3, 9, 11)

  • [Distinguishing existing methods and proposed methods? ] For our conditional stochastic compositional optimization problem, we have developed a new algorithm, CM-SCSC, as shown in the revised Section 5.2. Compared with the SCSC algorithm we used previously, the proposed CM-SCSC is better suited to our problem, where the inner expectation also depends on all outer random variables. We have further provided its theoretical analysis (i.e., the convergence results in Theorem 6 with proof in Appendix H.5) and revised numerical results (in Section 6 and Appendix G). For the EIG method in Appendix D.2.2, we have clarified that it is based on existing literature, to avoid misleading readers into thinking it is a new method proposed by us.

  • [Which experimental losses satisfy Assumption 2? How do the experiments satisfy the convergence assumptions?] In the revised Appendix F, we provide detailed definitions of the training losses and verify the conditions required for the convergence analysis of CM-SCSC in all three applications.

    For the newsvendor and two-stage inventory substitution problems, the training losses satisfy Assumption 2 in Section 5.1, including the required boundedness, Lipschitz continuity, and smoothness with respect to the decision-rule parameters. We also report the approximation errors introduced by the truncation method. For the portfolio problem, where Assumption 2(I) may not be satisfied when the perturbed returns have unbounded support, we introduce an alternative Condition 2 and verify that it holds for our portfolio loss. Together with Assumption 2(II), this condition is sufficient to establish the same convergence result for CM-SCSC. Finally, in our experiments, the step lengths and batch sizes follow Theorem 6, thereby ensuring that the convergence result for CM-SCSC applies.

  • [How is the p=1 portfolio objective kept finite?] In Appendix F of the revised manuscript, we defined the training loss of the portfolio, denoted by Ψ~Portfolio\widetilde{\Psi}_{\text{Portfolio}}, which is bounded on [0,B][0, B]. Then, for any feasible decision and uncertain return, we have 1≤EWϵ[exp⁡(Ψ~Portfolio/λϵ)]≤eB/λϵ<∞1 \le \mathbb{E}_{W_{\epsilon}}\left[ \exp\left( \widetilde{\Psi}_{\text{Portfolio}} / \lambda\epsilon \right) \right] \le e^{B/\lambda\epsilon} < \infty, where WϵW_{\epsilon} is a Laplace distribution with unbounded support. This also ensures that the outer exponential expectation with respect to UϵU_\epsilon is finite. Therefore, the modified portfolio loss is well-defined for p=1p=1 without truncating returns or perturbations.

  • [Comparison with smoothing baselines? ] We have added a smoothing baseline, i.e., a sample average approximation formulation with a kernel-smoothed probability distribution, and further compared it with the proposed Causal-SDRO method. The results show that the smoothing baseline performs well, while additional distributional robustness may yield further improvements. See the resulting model and numerical results in Appendix G.4.

I thank the authors for the detailed reply. My concerns have been addressed.

We thank the reviewer for the positive and constructive feedback.

Reviewer_6td7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies contextual distributionally robust optimization with causal transport constraints and absolutely continuous adversarial distributions. It introduces a Causal Sinkhorn Discrepancy, derives a dual formulation and a Gibbs-mixture characterization of the worst-case distribution, and uses a differentiable Soft Regression Forest (SRF) as the decision rule. A stochastic compositional algorithm is proposed with a claimed stationarity complexity of O(ε^(-4)). Experiments cover synthetic newsvendor and inventory substitution problems and a portfolio task using real market data.

I appreciate the motivation and the effort to connect modeling, interpretable policies, and applications. However, the current version does not establish a nontrivial role for the causal constraint in the reported experiments, and the correspondence between the algorithm, proof, and experimental objective remains unresolved.

Critical strengths
  1. Meaningful problem and coherent framework. The paper connects information-constrained transport with entropy regularization, policy learning, and optimization. If the issues below are resolved, this could provide a useful framework for structured contextual DRO.
  2. Breadth of evaluation. The experiments span single-stage and two-stage decisions and include a rolling-window portfolio evaluation on real data. The reported SRF results and interpretability analysis provide initial evidence of practical potential.
Critical weaknesses
  1. The causal constraint appears vacuous in the continuous-feature experiments. Section 5.1 defines the empirical conditional distribution through exact feature matching (p. 7): I(x ̂) = {j:x ̂_j=x ̂} Appendix F.1 samples features from a continuous Gaussian distribution, and F.2 uses the same construction (pp. 25 and 28). Unless an unreported grouping or discretization step is applied, the observed feature vectors are almost surely distinct. The empirical outcome is then deterministic given the empirical feature, so the constraint X ⟂Y ̂|X ̂ holds automatically for every coupling. With identical costs, reference measures, and regularization parameters, Causal-SDRO and SDRO should therefore have the same objective. The reported performance differences cannot, by themselves, establish a benefit from causal restrictions (Fig. 3; Appendix C.1). Please clarify whether these differences arise from conditional estimation, optimization, or other implementation choices. This directly affects the evidence for the main modeling contribution.
  2. The algorithm, convergence proof, and experimental objectives are not fully aligned. Algorithm 1 maintains one scalar and one finite-dimensional vector (p. 8). In contrast, Appendix G.4 introduces condition-dependent states: y_2^k (x ̂,ξ_1),y_1^k (x ̂) Equations (45)–(46) update these states for every condition and add clipping absent from the pseudocode (pp. 43–44). For continuous perturbations, such function-valued states are not equivalent to a single vector. The paper needs a finite implementation and a proof for that implementation, including its memory and per-iteration costs. As written, Theorem 6 does not establish the claimed guarantee for Algorithm 1. There is also a concrete integrability issue. The portfolio loss is quadratic (Appendix F.3.1, p. 30). Under the untruncated Laplace-type kernel in Eq. (4), the following expectation diverges for p = 1 and any finite positive penalty parameter: E_(ξ₂ ∼ Wε) [exp((Ψ(f_θ (x ̂+ξ_1),y ̂+ξ_2))/λε)] = ∞ The quadratic term dominates the linear tail penalty for any nonzero portfolio. Thus, the stated unbounded-support model does not yield a finite 1-Causal-SDRO portfolio objective. If the implementation truncates the support or modifies the loss, those choices and the corresponding objective must be specified. A finite Monte Carlo estimate alone does not establish integrability. Together, these issues prevent me from verifying that the experiments and complexity claim apply to the stated model.
Questions for authors

These are clarification requests whose answers could materially change my assessment.

  1. How are empirical conditional distributions constructed? Please report the numbers of observations per distinct feature vector in the synthetic experiments and any grouping or smoothing. If each feature occurs once, how do you explain the Causal-SDRO versus SDRO gap? A comparison using identical parameters and a common solver would help resolve this.
  2. Does the implementation maintain the scalar/vector states in Algorithm 1 or the conditional functions in Eqs. (45)–(46)? For the latter, how are continuous conditions represented and updated with finite computation? Please specify the memory cost, per-iteration cost, and clipping used.
  3. How is the portfolio objective made finite for p = 1? Are returns or perturbations truncated? If so, please provide the actual support, sampling distribution, normalization, and revised objective, and explain which theoretical results cover that implementation.
3 comments

We very much appreciate your review comments. The following provides point-to-point responses, and the revised parts of our manuscript are highlighted in blue\textcolor{blue}{\text{blue}}.

  • [How are empirical conditional distributions constructed?] In revision, we clarify the number of uncertain parameters associated with each covariate, i.e., nm∈1,5,10n_m \in \\{1, 5, 10\\}, in synthetic experiments (newsvendor and inventory substitution problems). For the portfolio problem with real-world data, we employ a k-nearest neighbors (k-NN) approach to group similar covariates, each group containing nm∈1,10n_m\in\\{1,10\\} observations.

It is true that when only one uncertain parameter is associated with one covariate, Causal-SDRO and SDRO are equivalent. In our previous experiment results, the observed performance discrepancy between Causal-SDRO and SDRO was due to differences in the computational algorithms and hyperparameter choices. In the revised experiments, we use identical hyperparameters for Causal-SDRO and SDRO. Moreover, since the two formulations are equivalent when nm=1n_m=1, we only solve Causal-SDRO and use the resulting solution for both methods, ensuring identical numerical results in this case. See the Table in our general response to the whole reviewer team for the example of newsvendor, and detailed box plots for other experiments are provided in the manuscript (Figures 3, 9, 11)

  • [How does the proposed algorithm work?] Thanks for pointing out the inconsistency between our algorithm and our theoretical analysis. In the revision (Section 5.2), we developed a new algorithm for our problem, termed CM-SCSC, that maintains vector states to track the nested expectation terms. Specifically, at each iteration, we draw QQ i.i.d. samples of ξ2\boldsymbol{\xi}_2 from WϵW_\epsilon, to estimate the inner expectation with respect to ξ2\boldsymbol{\xi}_2 conditional on each historical covariate value x^m\widehat{\boldsymbol{x}}_m and sampled ξ1\boldsymbol{\xi}_1. Each estimate is an NN-dimensional vector, where NN is the number of all historical observations. The outer expectation of ξ1\boldsymbol{\xi}_1 conditional on x^m\widehat{\boldsymbol{x}}_m is tracked by a scalar variable for each covariate group. These scalar variables are stored as an MM-dimensional vector, where MM is the number of distinct observed covariate values. We have also corrected the theoretical analysis accordingly (See Theorem 6 with proof in Appendix H.5).

    In the pseudocode for CM-SCSC, we clarify the clipping used in tracking variable updates. Remark 7 further shows that for a target convergence threshold ε\varepsilon, the CM-SCSC's per-iteration computational cost (the number of function evaluations in each iteration), per-iteration gradient complexity (the number of gradient computations in each iteration), and memory cost are O(ε−1)\mathcal{O}(\varepsilon^{-1}), O(ε−1)\mathcal{O}(\varepsilon^{-1}), and O(1)\mathcal{O}(1), respectively. We have also re-conducted the experiments using the revised algorithm, and they demonstrate that our model and algorithm continue to exhibit superior performance (See Section 6 and Appendix G).

  • [How to avoid infinite objectives of portfolio with p=1p=1?] In Appendix F, we provide detailed explanations of losses for all three applications. In particular, the loss for the portfolio problem is truncated and smoothed so that our algorithm's convergence result (Theorem 6) holds. Since the loss of the portfolio, denoted by Ψ~Portfolio\widetilde{\Psi}_{\text{Portfolio}}, is bounded on [0,B][0, B], for any feasible decision and uncertain return, it holds that 1≤EWϵ[exp⁡(Ψ~Portfolio/λϵ)]≤eB/λϵ<∞,1 \le \mathbb{E}_{W_{\epsilon}}\left[ \exp\left( \widetilde{\Psi}_{\text{Portfolio}} / \lambda\epsilon \right) \right] \le e^{B/\lambda\epsilon} < \infty, where WϵW_{\epsilon} is a Laplace distribution with unbounded support. This also ensures that the outer exponential expectation with respect to UϵU_\epsilon is finite. Therefore, the modified portfolio loss is well-defined for p=1p=1.

Dear Reviewer 6td7,

Thank you very much for your constructive feedback. We have incorporated the comments from the entire reviewer team into our revised manuscript and provided our responses. We would be grateful to know whether these responses address your concerns. We greatly appreciate any further feedback from you. Thank you for your time.

Dear Reviewer 6td7,

Thank you again for your feedback. We would appreciate it if you can look our response and let us know whether our clarifications address your concerns. We welcome further discussion and would be happy to answer any remaining questions.

Thank you for your time and consideration.

Best Regards, Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Overall Response to Whole Reviewer Team

We have incorporated the reviewer team's constructive feedback to revise our manuscript. The revised parts of our manuscript are highlighted in blue\textcolor{blue}{\text{blue}}. We summarize our major revision points below.

  • [How are empirical conditional distributions constructed?] In revision, we clarify the number of uncertain parameters associated with each covariate, i.e., nm∈1,5,10n_m \in \\{1, 5, 10\\}, in synthetic experiments (newsvendor and inventory substitution problems). For the portfolio problem with real-world data, we employ a k-nearest neighbors (k-NN) approach to group similar covariates, each group containing nm∈1,10n_m\in\\{1,10\\} observations. The table below provides the median out-of-sample performance improvement (from 10 independent trials) for the newsvendor problem with different sample sizes NN, dimensions dxd_x, and nmn_m. Detailed box plots for other experiments are provided in the manuscript (Figures 3, 9, 11). From these results, we find that as long as nm>1n_m>1, Casual-SDRO consistently outperforms SDRO. Other added experiment results are provided in Appendix G.
ModelNN(nm,dx)=(1,5)(n_m,d_x)=(1,5)(nm,dx)=(5,5)(n_m,d_x)=(5,5)(nm,dx)=(10,5)(n_m,d_x)=(10,5)(nm,dx)=(1,10)(n_m,d_x)=(1,10)(nm,dx)=(5,10)(n_m,d_x)=(5,10)(nm,dx)=(10,10)(n_m,d_x)=(10,10)
1-Causal-SDRO2069.0549.8434.8154.4337.4419.32
4068.8955.8445.4754.2252.4539.68
8078.0841.0956.4353.7237.8646.65
10074.4853.6153.9952.8338.6840.67
1-SDRO2069.0535.438.3254.4320.88-13.65
4068.8935.3430.3954.2237.6612.35
8078.0836.3540.8853.7212.7825.63
10074.4837.2249.8752.8325.9429.86
2-Causal-SDRO2071.5741.6422.4751.5131.617.94
4073.0048.1339.0049.2445.3835.71
8076.7150.5154.9755.4141.5339.96
10076.7052.8257.2054.3736.6940.36
2-SDRO2071.5729.005.9851.5115.73-27.45
4073.0038.9033.0949.2438.3617.54
8076.7132.7847.6355.4125.3528.13
10076.7050.8946.6654.3732.9932.25
  • [Algorithm Framework and Experiment Setup?] We follow reviewers' comments to revise the algorithm framework and its convergence to align with our experimental implementation (See Section 5 and proof in Appendix H.5). In Appendix F of the revised manuscript, we provide detailed explanations of losses for all three applications. In particular, the loss for the portfolio problem is truncated and smoothed so that our algorithm's convergence result (Theorem 6) holds.

  • [Comparison with smoothing baselines? ] We have added a smoothing baseline, i.e., a sample average approximation formulation with a kernel-smoothed probability distribution, and further compared it with the proposed Causal-SDRO method. The results show that the smoothing baseline performs well, while additional distributional robustness may yield further improvements. Please see the resulting model and numerical results in Appendix G.4.

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

We thank the AC for helping send a reminder to the reviewers to engage in discussion. Thank you for your help and time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.