Conformal Inverse Optimization

Inverse optimization has been increasingly used to estimate unknown parameters in an optimization model based on decision data. We show that such a point estimation is insufficient in a prescriptive setting where the estimated parameters are used to prescribe new decisions. The prescribed decisions may be low-quality and misaligned with human intuition and thus are unlikely to be adopted. To tackle this challenge, we propose conformal inverse optimization, which seeks to learn an uncertainty set for the unknown parameters and then solve a robust optimization model to prescribe new decisions. Under mild assumptions, we show that our method enjoys provable guarantees on solution quality, as evaluated using both the ground-truth parameters and the decision maker's perception of the unknown parameters. Our method demonstrates strong empirical performance compared to classic inverse optimization.

Paper

Similar papers

Peer review

Reviewer Bo7T7/10 · confidence 2/52024-07-10

Summary

The paper proposes a robust optimisation method for decision problems that provides not only a point estimate but also an uncertainty set. The methodology is tested on decision making tasks such as shortest path and the knapsack problem and shows favourable results when compared to classical inverse optimisation methods. The theoretical results focus on the validity of the uncertainty set as well as bounds on actual and perceived optimality of the decision policy.

Strengths

The paper addresses an important problem of inverse optimisation taking advantage of the CP framework. The proposed pipeline is clear and easy to follow. To the best of my knowledge, the work is original. The writing quality and clarity is good.

Weaknesses

My only concern is the practical applicability of the methodology. While I appreciate the motivation for the approach and the extensive discussion of the evaluation metrics, I feel the gains offered by the methodology are minor, particularly given the fact that large datasets will be needed to calibrate the model and choose the hyperparameters.

Questions

Can you explain the significance of the $\Delta_\alpha$ term in more detail (line 258)? While you mention some inverse reinforcement learning literature (e.g. line 72), I wonder if you are aware of any similar probabilistic approaches in that literature? I appreciate that providing a real-world example may be difficult, but can you comment on the possible issues of using your approach on real-datasets. Are any of the assumptions in your analysis likely to be violated (e.g. due to distribution shift or quality/size of data)?

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations are discussed in Sec. 6 as well as throughout the analysis and experiment sections.

Reviewer ggLM7/10 · confidence 2/52024-07-12

Summary

This paper proposes conformal inverse optimization (CIO) which uses conformal prediction to learn uncertainty sets for parameter estimates in IO. These uncertainty sets are then used to formulate a robust version of the forward optimization problem. The authors provide intuition for their method and also provide theory to show that CIO is a principled approach. Experiments on synthetic shortest path and knapsack problems demonstrate the promise of CIO.

Strengths

* The authors motivate their approach of conformal inverse optimization using examples (e.g., Example 1). * The figures in section 3 provide helpful intuition for the proposed approach. Nicely done! * The authors show that their method is theoretically justified — it obtains uncertainty set validity and also allows for a bound on the POG. * The authors provide a code implementation of their method, which makes their experiments easier to reproduce. Furthermore, the experiments are described in sufficient detail in the main paper.

Weaknesses

* The setup in section 3.1 could benefit from more intuition. I think having a running example (e.g., rider-driver matching) where $x$, $u$, $\theta$, and $\mathcal X(u)$ have a practical meaning would be very helpful. * Line 85: I think the domain of $f$ should be $\mathbb{R}^{n} \times \mathbb{R}^{d}$, not $\mathbb{R}^{n \times d}$. * The idea of $\alpha$ (line 202) could be explained better. How exactly does $\alpha$ control the size of the uncertainty set?

Questions

* Line 124: What is meant by “algorithm aversion”? * How is the uncertainty set calibration related to the classic conformal prediction framework?

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

Please see “Weaknesses”

Reviewer UvTc7/10 · confidence 3/52024-07-12

Summary

This paper considers a robust inverse optimization problem. The author provides an uncertainty set with probabilistic guarantees for the estimated parameters, illustrates the advantages of robust inverse optimization over deterministic inverse optimization through an example, and demonstrates its superiority in experiments on other problems.

Strengths

- This paper considers robust inverse optimization problems, which is a good complement to existing inverse optimization literature. - The procedure of efficiently solving CP is described in good detail, which is helpful for implementation. - The paper is overall clearly written and well-structured. The authors use a simple intuitive example to illustrate the motivation of robustifying the inverse optimization, which I find really helpful.

Weaknesses

- The present of the uncertainty set in (9) may need justifications. - Although the RFO method can work based on any point estimation method, the performance of the solution usually depends on the quality of the point estimation method. I wonder if the authors could discuss this in the experiments.

Questions

- The failure of the point estimation method in Example 1 is partially because of the sub-optimality loss. Will the conclusion be changed if using another loss?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations in Section 5 & 6.

Reviewer 5FdS5/10 · confidence 3/52024-07-16

Summary

In some supervised decision making, one usually leverage solution of optimization problem sequentially for future decisions. The authors argue that a single point estimation might not be enough and rather suggest estimating a whole uncertainty set on the model parameter and instead suggest a minimax robust optimization problem for future decisions. A nice contribution is to provide theoretical guarantees on the optimality gap of the proposed solution, with high probability.

Strengths

- The problem considered is quite interesting for safer AI based decision making. - The derivation of uncertainty sets seems novel and quite intriguing.

Weaknesses

The presentation is quite convoluted and hard to follow at several points. See questions below.

Questions

1-) **Presentation** : After introducing the framework, the introduced example 1 seems quite toyish and the presentation here is quite confusing: - example 1 is followed by lemma 1 that only applies to example 1, which is followed by proposition 1 that only applies to example 1 ... So I am quite curious to see how this generalize to other problems. Otherwise remove the "lemma" and "proposition" denomination and set it all under the Example 1 umbrella. - Same for lemma 2. - I am curious how the main contributions Theorem 2 and 3 applies to the given example 1. Also how does it applies to classical linear model with Gaussian noise? - In line 86, it seems that the authors assume the knowledge of the distribution $\mathbb{P}_u$. But it is not clear how is this used in the given examples, and how does it appears in the main results. 2) **Related works** : The authors might want to *remove* some overclaim in the paper e.g. in the conclusion *"We present the first approach to learning uncertainty sets from decision data, which is then utilized in a robust optimization ..."*, here are some refs of similar previous work, just to name a few. - (Siddiqui et al 2015) https://link.springer.com/article/10.1057/jors.2014.41 - (Csaji etal 2015) https://arxiv.org/abs/1807.08216 - (Dekker etal 2008) https://hal.science/hal-03046711 - (Guille-Escuret etal 2024) https://arxiv.org/abs/2401.15254 Beyond the downstream applications that might differ, how the contributions of this paper compare to such previous work? 3-) In the main text, it is unclear how conformal prediction framework is used. As far as I know, uncertainty sets derived from conformal prediction applies to *observations* and not model parameters directly [besides some very recent works that provide early results for such extension]. 4-) How does one obtain a reasonable $\eta$ in assumption 2. How such a choice affects the presented bound? Typically I am surprised it does not even appears in Equation 12. For example if one consider overparametrized linear model, the sets of solutions are usually unbounded. I also do not understand why we can use scale invariance here without loss of generality. The authors might want to write this part more explicitly _______ Edit after rebuttal _________ After considering authors rebuttal and discussions, I raised my score to support paper acceptance.

Rating

5

Confidence

3

Soundness

2

Presentation

1

Contribution

2

Limitations

OK

Authorsrebuttal2024-08-12

We thank the reviewer for responding to our rebuttal. - **Regression.** We note that in this paper we do not observe $f(\theta, x)$, we only observe $x$. So, we can not apply existing methods to regress $f(\theta, x)$ against $x$. For example, consider FO to be a shortest path problem on a road network with $d$ road segments, our context $u$ specifies an origin and a destination of a trip, decision $x$ is a $d$-dimensional binary vector representing if each road segment is selected (=1) or not (=0), and $\theta$ is a $d$-dimensional real vector representing the (perceived) travel cost on each road segment. In this setting $f(\theta, x)$ is the perceived travel cost of a trip, which we never observe. We only observe the trip ($x$) and its origin and destination ($u$) (see line 99 for the data we are dealing with). We will add this running example in Section 3 to add clarity. Note that we had confusion about this comment in the initial rebuttal. We thought the reviewer meant we could regress $x$ against $u$. So, we use the term *decision* to highlight that $x$ is subject to constraints, so this regression idea is not applicable. - **Relation to CP.** Our result in equation (11) is a conformal prediction result. To build a conceptual connection, we can think of our observation $(u, x)$ as features, our inverse optimization problem (5) as an ML model, and the estimation target $\theta$ as the prediction target in CP. We apply the simplest CP method to construct a set (which is the same for all $(x, u)$) that contains a $\theta$ that can explain thee next $(u, x)$ with a specified probability. We hope this analogy helps to understand the connection between our work and CP. - **Distribution.** We apply concentration inequality to bound the difference between the *observed probability of even E* and the *real probability of event E*, where E indicates there exists a $\theta$ in our set that can explain the next $(u, x)$. Since *probability* is bounded in $[0, 1]$, we do not need any distributional assumptions. - **Ground-truth parameters**. Different data points are generated by different $\theta_k$ which are noisy perceptions of the ground-truth $\theta^*$. In the above shortest path example, $\theta^*$ is the ground-truth travel time on each road segment and $\theta_k$ are noisy human perception. As mentioned above, we will add this example to the paper to add clarity. - **Parameter $\eta$**. Yes, in short, our method can provide a smaller set than the bounded support set. This is very important in our setting because our uncertainty set will be used to prescribe future decisions. If the set is too large, then the solution will be too conservative. We point the reviewer to lines 202-204 where we mention we want the smallest set that can achieve the coverage requirement. Our bounds in Theorem 3 also decreases as the size of the set $\alpha$ decreases, reflecting this intuition. - **Selection $\eta$**. We'd first like to emphasize that we do not need to select a $\eta$ to apply our method, so the computational cost should not be a concern in practice. The purpose for calculating $\eta$ in Example 1 is simply to verify our theoretical results. Regarding the estimation method we use, as stated in Assumption 2, $\eta$ is the largest possible diameter of $\Theta(x, u)$. So calculating the diameter of $\Theta(x, u)$ for all possible $(x, u)$ should give as a valid $\eta$. In real problems, there might be a large number possible $(x, u)$, so calculating $\eta$ could be challenging. But again, we don't need to calculate this parameter in practice. We will add the description of how we calculate $\eta$ in Table 1 in the appendix and mention this in the main body when presenting Table 1. Thank you again for your questions which helped to improve the clarity of our paper. We'd be happy to discuss further if there are other questions.

Reviewer 5FdS2024-08-12

Thanks for the detailed clarifications. I think adding them in the paper would be helpful to the readers.

Authorsrebuttal2024-08-12

We will make the following changes in the final version according to the reviewer's comments. - Add explanations in Section 3 to clarify that our problem is far from a linear regression problem because - We do not observe $f(\theta, x)$ so we can not regress $f(\theta, x)$ against $x$ - Our observations $x$ are solutions to optimization models, i.e. $x = argmin_{x’\in X(u)} \sum_i \theta_i f_i(x)$, so we can not regress $x$ against $u$ using a simple linear regression model. We will also add the shortest path running example to add clarity. - Add explanations in Section 4 to describe the strong connection between our work and conformal prediction. We will emphasize - Theorem 2 is entirely distribution-free. - We extend traditional CP guarantees to the context where covariates are used to predict a set that must intersect with the set that is observed aposteriori. When the aposteriori set is known to be a singleton, our result reduces to classical CP results from Vovk et al. 2005. - Add explanations in Section 3 to emphasize - We do not assume any knowledge of $\eta$ and only assume its existence - Add discussions in Section 4.2 (when presenting Table 1) to talk about how we might calculate $\eta$ to validate our theoretical results. We thank the reviewer for the valuable feedback. We believe these changes will improve the clarity of our paper. In light of these changes, we would appreciate it if the reviewer could revisit the choice of score for our paper.

Reviewer UvTc2024-08-09

Thank the authors. All the points I was interested in have been adequately addressed. I have no further questions or concerns.

Authorsrebuttal2024-08-09

We thank the reviewer for the insightful comments and questions which helped improve our paper.

Reviewer Bo7T2024-08-10

I thank the authors for the response. My only remaining concern is relation to related literature (as highlighted by other reviewers as well). However, I am not aware of any specific papers myself, and so I will keep my original score.

Authorsrebuttal2024-08-10

We thank the reviewer for the response and the overall positive appreciation of our paper. During the rebuttal, we performed another round of literature review and still did not find any existing probabilistic approaches directly applicable to our setting. Therefore, we suggest that our discussion regarding the existing work is sufficiently comprehensive and appropriate. Regarding the literature on learning a confidence region for parameters in a linear regression model (as suggested by reviewer 5FdS), we point the reviewer to our rebuttal to the same reviewer related to **linear model** and **related work**. Below we highlight the main differences between this and our settings. - **Different mappings.** In our setting, we are given (context $u$, decision $x$) data and the mapping from $u$ to $x$ is described by an optimization layer parameterized by $\theta$. Our decision $x$ are subject to constraints. In regression, we are given (feature $u$, target $x$) data, the mapping is a closed-form linear function $x = \theta^\intercal u$ and $x$ is unconstrained. As existing regression approaches do not capture the optimization structure, they are not applicable in our setting. - **Data generation process.** Our data points $(u_k, x_k)$ are generated by different $\theta_k$, reflecting different decision makers may have different intuitions. In contrast, for regression problems, all data points are generated by the same $\theta$. - **Goal.** We are looking for an uncertainty set that contains parameters to explain the next unseen $(u_k, x_k)$, while the regression approaches are looking for a set that contains the true $\theta$. This leads to different theoretical analysis in our paper (Theorems 2 and 3). As mentioned in our response to reviewer 5FdS, we will add these discussions to Section 3.2 to highlight the different between our setting and the regression setting. We hope this explanation will address the reviewer's concern. If there are any further questions, we would be happy to discuss them in the following days.

Reviewer ggLM2024-08-11

Thanks for the rebuttal

Thank you for addressing my comments. I have raised my score accordingly.

Reviewer 5FdS2024-08-12

> We see the connection between linear regression and our setting, but since they are very different, it is unclear how to interpret the theorems in this new setting Regardless of interpretation, the paper consider $f(\theta, x) = \sum_i \theta_i f_i (x)$ to be a generalization of $f(\theta, x) = \theta^\top x$. However, since $f_i$-s are independent of $\theta$, one simply set $z_i = f_i(x)$ to equalize the two settings. So it is not clear how this changes anything for the results of the paper or how meaningful it is. In a sense, without loss of generality, it seems to me that one can simply consider linear models in its simplest form. The vocabulary change "context/feature" and "decision/target" seems irrelevant. Some more clarification might be needed. > Relation to conformal prediction Vovk etal 2005 I do not see any connexion between the presented work and CP from Vovk etal, it seems that no conformal prediction results have been used in this paper. So I am a bit confused by the title and narrative. The results of this paper does not seems distribution-free since they leverage concentrations inequalities which request more assumptions on $\mathbb{P}$ > different data points are generated by different parameters. Then there might be a source of confusion in the problem setup in line 84-91: "There exists **a ground-truth parameter $\theta^\star$**" > On the hyper parameter $\eta$ By definition, if you assume a knowledge of $\eta$ such that $\|\theta - \theta'\| \leq \eta$, then the problem is solved since this itself define a confidence set no? Or is it given such $\eta$ your pipeline provide a more refined set? > enumerate possible pairs and calculate the diameter of the corresponding set defined in equation (2). This is how we calculate the bounds in Table 1. Regarding its impact, based on equations (13) and (14), the bound increases as increases. Why this provides a provably valid selection of $\eta$? What is the additional computational cost? Also this seems to leads to higher uncertainty set no? More descriptions on this parameter could be helpful.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC