acceptodds
Under review as a conference paper at ICLR 2027

Population-Preserving Loss Corrections for Finite-Sample Decision Selection

Abstract

Losses with identical population comparisons can select different decisions from finite samples. We study shared loss corrections learned across tasks and deployed by empirical minimization over fixed candidate banks. Conditionally centered controls preserve population preferences while changing joint empirical comparisons. We characterize harmful selection through multivariate Chernoff bounds, an entropy dual, and a convex-reference support certificate for zero regret. Regularized exponential construction in a reproducing kernel Hilbert space has a finite positive-cone representation, but convex comparison criteria admit an unbounded multiplicative gap to decision risk, even for bounded losses and one correction parameter. Direct episodic welfare fitting over a -parameter affine family instead achieves excess risk of order from independent tasks with at most candidates, without a margin assumption. A centered sensor construction gives a matching lower bound for fixed bank size, even when each sampled task's law is revealed. Independent validation adds finite-library oracle and baseline-safety guarantees. Controlled sensor and contextual studies examine functional capacity, joint comparison information, and the difference between convex construction and direct decision-risk fitting.

Primary area:
learning theory
Keywords:
learning theory, decision-focused learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9CGyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers picking one of a fixed set of candidate predictors for a given task from a small sample of that task’s data. Rather than picking the predictor with the lowest sample loss as is, it proposes adding to the loss a learned control variate with zero mean, which can change the preferred predictor on a small sample. The paper develops theory on when such terms help and on learning them from past tasks, and illustrates it on simulated examples.

Critical strengths

The results appear correct.

Critical weaknesses

This paper offers nothing substantially new. It repackages textbook results for the selection rule and presents them as its theorems: a Chernoff bound, a representer theorem, a Jensen counterexample, the generalization bound of [1, Theorem 3.3], an Assouad lower bound and a union bound. The control-variate idea is already used in ranking and selection [2]. The experiments are all toy simulations.

[1] Balcan et al. How Much Data Is Sufficient to Learn High-Performing Algorithms? Generalization Guarantees for Data-Driven Algorithm Design. STOC, 2021.

[2] Tsai and Nelson. Combined Ranking and Selection with Control Variates. Winter Simulation Conference, 2006.

Questions for authors

None.

Reviewer_77f6OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Summary: The paper studies how to improve finite-sample selection from a fixed candidate bank by learning shared loss corrections across tasks. Its key finding is that optimizing convex comparison criteria can lead to arbitrarily large relative decision-risk gaps, despite population preservation. Direct episodic welfare fitting offers an alternative with provable learning guarantees, while independent validation supports selection with baseline-safety guarantees. Controlled experiments illustrate the distinction between population alignment, tractable convex construction, and actual finite-sample decision quality.

Critical strengths
  1. A useful insight into finite-sample decisions. Theorem 3 clearly shows that preserving population rankings does not guarantee good empirical choices—even optimizing a convex comparison criterion can lead to much higher decision risk. This is a compelling reason to evaluate corrections by their downstream performance.

  2. Strong guarantees for learning corrections directly. Direct episodic fitting has a margin-free learning guarantee and a matching lower bound for fixed bank size. Together, these results help explain how many independent source tasks are needed to learn a useful correction.

Critical weaknesses

Practical applicability needs more guidance. The main guarantees require exactly centered controls and independently fixed candidate banks, while efficient exact fitting is demonstrated for separable histogram corrections. A concrete application and clearer computational guidance would help readers understand where the approach is practical.

Questions for authors

NA

Reviewer_UNFtOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

The paper studies how to improve finite-sample decision selection without changing population loss comparisons. The authors introduce loss corrections using conditionally centered controls. They also develop many theoretical results. The paper discusses the limitations of convex comparison criteria and proposes direct episodic learning to minimize decision risk. Upper and lower bounds on excess risk are established. Numerical results are provided to evaluate the methods.

Critical strengths

The paper considers an important problem. Two losses can have the same population minimizes but different finite-sample performance. The authors study this problem through several theoretical results. I find Theorem 3 (p.5) interesting. It shows that a convex comparison criterion can select a correction with much larger decision risk. Theorems 4 and 5 (p.6) also establish upper and lower bounds for direct episodic learning. These results help us understand why minimizing a convex criterion may not lead to the best decision.

The numerical experiments also cover lots of settings. Section 6 (pp. 7-9) compares different correction methods and function classes. The authors discuss cases where direct episodic fitting performs worse than entropic fitting. For example, the results on p. 9 show that entropic fitting performs better with fewer source tasks in some settings. I appreciate that these limitations are reported.

Critical weaknesses

I have two main concerns.

First, I have some questions about the assumptions. In Section 4 (p. 4), the correction requires a control variable with conditional mean zero. It also requires a known population-preserving function space. I am not sure how these conditions can be checked in practice. Appendix H.2 (p. 39) considers approximate centering, but the experiments mainly use controls that are centered by construction. What happens when the conditional mean is unknown and must be estimated?

Second, I am concerned about the computational cost. Theorem 4 (p. 6) gives an excess-risk bound, but finding the empirical minimizer can be difficult. The experiments mainly consider correction families with relatively few parameters. It is not clear how the method would work for larger problems. I also have a question about Theorem 6 (p. 7). The safety gate returns the direct loss in every repetition of the main sensor experiment (p. 8). Similar results are reported in Appendix I.5 (p. 45), where all 80 gates return the baseline. Although the theoretical guarantee is valid, I am not sure how useful this procedure is in practice.

Questions for authors

My questions will be based on my previous comments for critical weaknesses.

(1) In Section 4 (p. 4), the method requires a control variable with conditional mean zero. How can we check this condition when the task distribution is unknown? I wonder how sensitive the method is to errors in estimating the conditional mean.

(2) Theorem 4 (p. 6) gives an excess-risk bound, but the optimization problem can be difficult to solve. How does the computational cost change when the number of parameters or candidate decisions increases?

(3) For Theorem 6 (p. 7), I am concerned about the safety gate. In several experiments, the gate returns the direct loss even when better corrections are available. How many validation tasks are needed before the gate can select an improved correction? I think this needs more discussion.

(4) All experiments use simulated data. These results help explain the theory, but I am not sure how well the method works with real data. A real-data experiment would strengthen the paper.

(5) Appendix I.3 (p. 43) mentions that "These supplementary experiments use banks of contextual predictors for combinatorial decisions. Their implementation is not included in content-E/experiments/." Why?

(6) The paper has so many theoretical results. I suggest authors to remove some of them (unless they are really critical to your method).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.