- 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 complexity for reaching an -stationary point. The method is evaluated on newsvendor, inventory-substitution, and portfolio problems, with comparisons across DRO models and policy classes.
Critical strengths
-
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.
-
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
-
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 is determined by under , so every coupling satisfies . Hence 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.
-
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 , and pooling the conditional expectations can change the gradient. Yet Algorithm 1 resamples and while maintaining one shared scalar and vector . The proof in Appendix G.4 instead maintains functions and for every condition (Eqs. (45)–(46)). Since 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 , 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?
- How does the implementation estimate the expectations separately for each ? Could the authors reconcile Algorithm
1 with Eqs.(45)--(46), including how the conditional estimates are represented and how their evaluation cost enters Theorem~6?
3 comments
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 (number of distinct observed covariate values) and (total number of historical observations). The number of iterations is and independent of . The per-iteration computational cost is .
-
We provide additional experimental studies on the newsvendor problem to examine the influence of and on the actual training time. In Table 1, we fix and increase by decreasing the number of observations associated with each covariate. In Table 2, we fix and increase by increasing the number of uncertain observations associated with each covariate. In Table 1, we find the training time increases with for fixed . However, the results in Table 2 indicate the training time remains relatively stable for for fixed . 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 .
Table 1. Training time (s) with fixed and increasing (averaged over 3 trials).
| 40 | 80 | 200 | 400 | 800 | 2000 | 4000 | |
|---|---|---|---|---|---|---|---|
| 1-Causal-SDRO Time (s) | 7.728 | 8.027 | 8.040 | 7.969 | 7.864 | 8.791 | 11.582 |
| 2-Causal-SDRO Time (s) | 7.131 | 7.516 | 7.139 | 7.242 | 7.806 | 8.137 | 10.956 |
Table 2. Training time (s) with fixed and increasing (averaged over 3 trials).
| 40 | 400 | 800 | 1600 | 2400 | 3200 | 4000 | |
|---|---|---|---|---|---|---|---|
| 1-Causal-SDRO Time (s) | 7.654 | 7.746 | 7.746 | 7.913 | 7.806 | 7.911 | 7.728 |
| 2-Causal-SDRO Time (s) | 6.749 | 7.114 | 6.896 | 7.159 | 7.176 | 7.171 | 7.131 |
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.
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=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)