acceptodds
Under review as a conference paper at ICLR 2027

Pareto-Optimal and Strategyproof Data-Sharing Between Competitors

Abstract

We study data-sharing among firms competing in the same market through the quality of their predictive models. We model firms' profit through two main components, precision and market share, where the former depends on the firm's own number of datapoints and the latter depends on the endowments of all firms. We characterize Pareto-optimal contracts as those in which at least one firm receives all of the other firms' datapoints. We then propose a parametric family of data-sharing contracts and show that, under any of them, the relative market shares of all firms remain constant, eliminating the negative effect of sharing data with competitors and equalizing all firms' gains relative to the no-sharing scenario. We further show that one contract within this family satisfies both Pareto-optimality and strategyproofness against firms attempting to under-declare the size of their initial endowment. Finally, using a real-world dataset, we empirically test the gain in profit for all firms under the contract described above, and show that data-sharing is particularly beneficial when the machine learning task is harder or when firms' initial endowments are closer to homogeneous.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
data-sharing, game theory, pareto optimality, fairness, strategyproofness

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

The paper studies data-sharing among firms competing on prediction quality. It characterizes Pareto-optimal sharing and proposes a contract that gives firms equal proportional profit gains. Under a concave learning curve, the contract discourages under-reporting of data holdings. Experiments use Expedia data to estimate learning curves and calculate potential profit gains.

Critical strengths

The equal-gain contract offers a clear solution to the tension between cooperation and competition. The Pareto characterization helps explain which sharing arrangements leave room for mutual improvement. The withholding analysis accounts for firms retaining their undisclosed data.

Critical weaknesses
  1. The strategy-proof guarantee applies to continuous allocations. Rounding can make under-reporting profitable: a firm may receive the same rounded allocation while reducing its competitor's allocation. Small average utility changes after rounding do not establish that truthful reporting remains optimal. Definition 9 also needs to reflect the retained private data used in the proof.

  2. The empirical claim of strategy proof is insufficiently supported. Monotonic interpolation does not establish the concavity required by the theorem. Some experimental holdings also fall below the stated calibration range. The reported profit gains depend on assumed customer behavior; the Expedia experiment measures prediction accuracy.

Questions for authors
  1. What incentive guarantee survives the actual rounding rule, and can Definition 9 be made consistent with the proof?

  2. Does the fitted learning curve satisfy the theorem's concavity assumption, and how are holdings below the calibration range handled?

3 comments

Thank you for your comments and remarks. Please find our answers below.

First, regarding the facts that strategyproofness does not exactly hold with rounded allocations and that the fitted precision functions are not entirely concave, please refer to points (1) and (2) of the answer to reviewer CfSG, where we respectively explain a new result on bounding the utility gain from under-reporting and rounding, and discuss the concavity of the fitted functions.

1. Correcting Definition 9. Regarding the mistake in Definition 9, we thank the reviewer for spotting this typo. We have corrected the definition in the revised article. Having in mind that, given a declared initial endowment n^\mathbf{\hat{n}}, Q(n^)∈R+m\mathbf{Q}(\mathbf{\hat{n}}) \in \mathbb{R}_+^m denotes the amounts of additional data received by each firm under a contract Q\mathbf{Q}, the new version of Definition 9 reads :

Let Γ\Gamma be a data-sharing instance. A data-sharing contract Q\mathbf{Q} is strategyproof if for any i∈Fi \in \mathrm{F} and any declared initial endowment n^\mathbf{\hat{n}}, it holds, ui(n+Q(ni,n^−i))≥ui(n+Q(n^i,n^−i)),u_i\big( \mathbf{n} + \mathbf{Q}( n_i, \hat{n}_{-i}) \big) \geq u_i\big( \mathbf{n} + \mathbf{Q}( \hat{n}_i, \hat{n}_{-i}) \big), where n^−i:=(n^j)j∈F∖{i}\hat{n}_{-i} := (\hat{n}_j)_{j\in\mathrm{F}\setminus\{i\}}.

Still, we remark that the proof of strategyproofness already used the correct definition so the original article presented no issue regarding the correctness of the result.

2. Extrapolation for holdings below the calibration range. We thank the reviewer for pointing out that some experimental holdings fall below the calibration range; after a thorough check, it happens only for three points, corresponding to the three values of kk for the pair (δ,m)=(0.6,10)(\delta, m) = (0.6, 10). For this pair, the smallest point nin_i is equal to 81; for these three points (one for each value of kk), the values were simply extrapolated automatically using the interpolation method (piecewise cubic Hermite spline) and extrapolating based on the nearest edge interval (i.e. the interval starting above 100). In total, extrapolation was therefore required in 3 out of 81 points computed for Figure 2.

Thank you for correcting Definition 9 and explaining the extrapolation below the calibration range. These answers clarify the treatment of retained private data and the below-range holdings. Please include the extrapolation procedure in §5.2.

The new Proposition 6 provides a useful answer to my rounding concern under concavity. Please state that assumption explicitly and distinguish the bounded gain from under-reporting from exact strategyproofness.

My remaining concern is the applicability of the incentive guarantee to the fitted precision functions. The concavity analysis you refer to documents violations, whereas §5.2, lines 365–368, still states that the assumptions of Theorem 5 are satisfied. Decreasing secant slopes at larger windows do not establish concavity over the full domain. Please reconcile these statements and clarify which incentive guarantee, if any, applies to the fitted-and-rounded implementation. For now, I maintain my weak-rejection recommendation.

Thank you for your additional remarks. Please find below our supplementary answers.

Regarding the applicability of the incentive guarantee to the fitted precision functions, please refer to the answer to reviewer CfSG and to the new Section 5.4 of the revised paper, where we provide an empirical calculation of the maximum gain from under-reporting in our experimental setting.

We remark that the new included experiment required an extrapolation of the function PP between 00 and 100100 datapoints. In order to obtain these values, we have changed the method explained in our previous answer by fixing P(0)=0P(0) = 0 and simply performing the interpolation on [0,100][0, 100] as we did in the full calibration range [100,20000][100, 20000]. Intuitively, P(0)=0P(0) = 0 represents the fact that a firm with no training data never has any of its offers accepted as such a firm has no way of training an ML model. We included this procedure as a footnote in §5.2.

Finally, the revised article presents a corrected version of Proposition 6 with concavity stated as a required assumption.

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

The paper studies a setting where mm firms compete on same market. Their goal is to offer a incoming customer an offer that will maximize their expected profit which is defined as the product of two quantities : precision and market share. Although each firm has its own dataset made off historical data of past clients, they may wish to acquire extra data points from other firms as those increase its precision and market share. The possible data exchanges between firms are captured through a contract Q: a matrix with real numbers, the goal of the paper is to study under realistic conditions what contracts will be realized. First they show that there exists a contract where at least one firm cooperates with everyone and that is pareto optimal; every company that participates will have gains at least as good as not participating, and the firm that cooperated with everyone will have strictly better gains. However, they take into account the fact that companies will likely not participate in a contract if they don't have equal gains with the other participants. Hence, they also characterize the family of contracts that will induce equals gains for all firms. Finally they take into account the fact that some companies may strategically undereport the amount of data they have, and they show that when the precision function is concave then firms will have no benefit to misreport.

Critical strengths
  • The paper is generally well written: clear motivation for the setup, clear definitions and sufficient explanations for results.
  • The central findings even given the assumptions, are interesting.
  • They offer a honest and useful discussion of their results and limitations.
Critical weaknesses
  • It is not clear for me what the precision function represents; it represents how well can the firm estimate the value of its data points for predicting the best offer for a new customer? In any case, more importantly why we assume all firms know it and it is common across them? Even in the experiments, it is computed from all the points that are then splitted into firms datasets; but the firm alone did not have access to those points.
Questions for authors

Minor:

  • In proposition 3, LiγL_i^{\gamma} is a real number, but we refer to LγL^\gamma as contract, what does LiγL_i^{\gamma} represent for each firm?
3 comments

Thank you for your comments and remarks. Please find our answers below.

1. Meaning of the precision function. Firstly, thank you for pointing out that the definition of the precision function is not clear. We meant to define the precision function as the probability that, upon the arrival of a client to a firm, the client will accept an offer from that firm. More precisely, P(n)P(n) is the probability that a client, when going to a firm using a model trained on nn data points, accepts the offer made by that firm. We have added this discussion in the paragraph prior to Definition 2.

2. Assumption that PP is common to all firms. We assume that PP is common across all firms because we implicitly assume that, if two firms have the same amount of training data points, they will build two models that will have the same performance. This is of course a simplifying assumption, but it stems from the idea that the best performing ML models are usually known by players in the market, and that these players will choose to use a model that reaches the best achievable performance, given the current state of the art. In particular, we assume that different firms will not have significant differences in model precision if they have the same amount of available training data. In economic terms, this would correspond to assuming perfect information (on the performance of available ML models) and homogeneous production functions.

3. Assumption that PP is known. Assuming that PP is known to firms is also undeniably a simplifying assumption required to derive analytical results; however, it can also be argued that every firm (with an amount of training data nin_i) would have a good approximation for P(ni)P(n_i), simply as its empirical success rate (the percentage of their past clients that accepted an offer). Then, combining these estimated values already leads to a rich set of good estimations of the function PP, that could then be completed by interpolation and extrapolation.

4. The precision function in the experiments. During the experiments, as you correctly underlined, we compute the function PP using all points available in our dataset; and we then assume that firms had access to disjoint sets of points, which is why we enforce the sum of all nin_i to always equal 2000020000. Notice that this is done since what matters in our framework is the total number of datapoints that several firms can obtain when sharing (which can go from the minimum nin_i up to the sum of all (n1,...,nm)(n_1,...,n_m).). In particular, the computation of the function PP and the distribution of datapoints between the firms are actually independent processes.

5. Clarifying LγL^\gamma and LiγL^\gamma_i. Regarding your question about the clarification between LγL^\gamma and LiγL_i^\gamma; please note that Lγ:=(L1γ,…,Lmγ)L^\gamma := (L^\gamma_1, \dots, L^\gamma_m) is a vector of size mm, where each LiγL_i^\gamma denotes the amount of additional data received by firm ii under contract LγL^\gamma. Here, we use the abuse of notation described at the end of Section 3, treating contracts and final endowments as equivalent. Still, we recognize that LγL^\gamma is not properly defined in the article, and so we have modified Proposition 3 to make it explicit.

Thank you for your comments. I think discussing more the precision and PP is worthy to motivate your setting. I will maintain my positive score.

Thank you for your quick reply. As suggested, we have added the discussion on the function PP to the revised version of the article.

Reviewer_CfSGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies data sharing between competing firms and proposes a contract that gives them the same proportional profit gain. It characterizes the Pareto frontier and proves strategyproofness against under-reporting in a continuous model with a common concave precision function. Expedia data are used to illustrate the resulting gains.

Critical strengths
  1. The paper captures a useful tension: sharing data can improve a firm’s own model while also strengthening its competitors. The characterization of the Pareto frontier gives a clear answer to part of this problem.

  2. I found the proof of Theorem 5 the most interesting part of the paper. It accounts for firms retaining their undeclared data and allows other firms to under-report as well. This is an important part of the incentive problem.

Critical weaknesses
  1. I am not sure the strategyproofness guarantee carries over to the implemented mechanism. Theorem 5 is proved for continuous allocations, while Section 5 rounds them. A small average change in utility does not tell us whether rounding makes under-reporting profitable. For example, take P(x)=x/(x+1)P(x)=x/(x+1), r=1r=1, P∅=0P_{\varnothing}=0, and true endowments n=(2,5)n=(2,5). Under either floor or nearest-integer rounding, truthful reporting gives final endowments (2,7)(2,7). If the first firm reports only one datapoint, these become (2,6)(2,6), increasing its utility from 32/11132/111 to 7/247/24. The endowments here are very small, so this example does not establish the practical impact. A bound on the maximum gain after rounding would help show whether this matters at the scales considered in the paper.

  2. Some of the empirical conclusions go further than the results support. Section 5.2 treats the fitted learning curves as establishing strategyproofness and robustness. However, the interpolation is described as preserving monotonicity, and the required concavity is not verified. It is also unclear how uncertainty in the estimated precision function affects the contract. Figure 2 needs a clearer interpretation. With P∅=0P_{\varnothing}=0, the common profit multiplier is simply P(20000)/P(n1)P(20000)/P(n_1), so these are gains calculated within the assumed market model. The explanation that increasing the number of firms makes their endowments more homogeneous also needs correction: under the proposed construction, n1/nm=δ−(m−1)n_1/n_m=\delta^{-(m-1)} increases with the number of firms.

Questions for authors
  1. What guarantee holds after rounding? Could you bound the maximum gain from under-reporting, or give an integer allocation rule that preserves strategyproofness?

  2. Is the fitted precision function concave over the full range used by the contract? How sensitive are the allocations and incentives to uncertainty in this function?

5 comments

Thank you for your comments and remarks. Please find below our answers.

1. Bound on the maximum gain from under-reporting. Regarding the remark on strategyproofness and rounded allocations, you are correct. Following your suggestion, we have deduced an upper bound on the utility gain that a firm can obtain by under-reporting under rounded allocations, that we added on the revised version of the paper as Proposition 6 and whose proof is included in Appendix B. Denoting nmin⁡n_{\min} the minimum amount of initial data across all firms, the gain of lying under rounded allocations, as a percentage of the utility obtained when declaring truthfully, is bounded by (P(nmin⁡)P(nmin⁡−1))4r+2\left( \frac{P(n_{\min})}{P(n_{\min} - 1)} \right)^{4r + 2}

Notice that, whenever PP is concave and nmin⁡→∞n_{\min}\to\infty, the bound goes to 11, implying a null profit from under-reporting. Additionally, in non-asymptotic regimes, small values of nmin⁡n_{\min} already guarantee a negligible profit from under-reporting. For example, considering P(x)=x/(x+1)P(x) = x / (x+1) and nmin⁡=100n_{\min} = 100, the maximum gain from under-reporting does not exceed 0.06%0.06\%, while nmin⁡=25n_{\min} = 25 already guarantees a gain no higher than 1%1\%.

2. Concavity of the fitted precision functions. Regarding the concavity of the empirical function PP, as you correctly remarked, upon closer inspection, the function does not happen to present a perfect concave behavior, as for example when the set of optimal hyperparameters changes (represented by the squares, triangles, and circles in Figure 1). To test the concavity of the empirical precision functions, we have analyzed whether the secant slopes decrease along the function domain, which would characterize a concave function. At each sampled point of the curve, we calculated the secant slope mi=yi−yi−sxi−xi−sm_i = \frac{y_i - y_{i-s}}{x_i - x_{i-s}} for a step size s∈{1,2,3}s \in \{1, 2, 3\}, and calculated the percentage of points for which mi≤mi−sm_i \leq m_{i-s}. The results are displayed in the table below. The corresponding discussion has been also included in Appendix C in the revised version of the paper.

Menu SizeStep s=1s = 1Step s=2s = 2Step s=3s = 3
k=1k = 166.67% (12/18)87.50% (14/16)100.00% (14/14)
k=3k = 372.22% (13/18)87.50% (14/16)100.00% (14/14)
k=5k = 583.33% (15/18)100.00% (16/16)100.00% (14/14)

For a step-size of s=1s=1, the secant slopes are not always decreasing, which implies that the functions are not entirely concave at a micro level, which can be attributed to the changes in hyperparameters. We expect that performing a more thorough hyperparameter selection would lead to smoother transitions, reducing the non-concave behavior. Additionally, with a larger window at s=3s = 3, the secant slopes are indeed all decreasing for all 3 fitted precision functions, which supports the claim that, at a macro level, the functions do display a concave behavior. Nevertheless, we have also modified the discussion about the concavity holding in Section 5.2 to a more nuanced version.

3. Sensitivity of the results to uncertainty in PP. Regarding the relation between the uncertainty of the function PP and the utility gains under contract HH from Figure 2, the revised paper presents now a stochastic version of the originally showed experiment where we have repeatedly computed the gains of firm 11 by sampling the function PP from the intervals obtained in Figure 11. A precise description is included in the revised paper between the lines 406 and 412. We observe no significant changes to the conclusions already presented in the submission. Even when considering the uncertainty on the function PP, the firms benefit from sharing data, particularly when the machine learning task is harder or the initial endowments are closer to homogeneous.

4. Impact of the number of firms on homogeneity. Finally, we thank the reviewer for spotting the mistake regarding n1n_1 and nmn_m being farther away as mm increases. We have modified the article to precise that, as the total number of datapoints nmax⁡n_{\max} is fixed, increasing mm actually creates a more homogeneous setting as more and more firms, the ones in the middle, obtain closer number of datapoints, even though the extreme ones get farther apart.

Thank you for adding Proposition 6. The bound is a useful response to my rounding concern. I have three remaining questions.

  1. Could you state the concavity assumption explicitly in Proposition 6 and confirm that the rounding rule used in the experiments satisfies its error condition?

  2. Since the fitted curves are not concave over the full range, what incentive guarantee applies to the actual implementation? Please clarify this and adjust the experimental claims accordingly.

  3. Could you report the largest gain from under-reporting in the experimental settings? The uncertainty experiment examines changes in the gains from sharing data, but does not directly answer whether under-reporting is profitable.

I appreciate the theoretical contribution and the additional analysis. For now, I am keeping my score at 2 because the connection between the new bound and the experimental implementation remains unclear.

Thank you for your additional remarks. Please find below our supplementary answers.

  1. Regarding the concavity requirement on Proposition 6, we have added the assumption to the statement. Also, we have simplified the statement regarding the rounding of the contract H. Thank you for pointing out this oversight.

  2. Regarding the concavity assumption on the experimental section, as correctly remarked but the reviewer, we have revised the article and stated that the empirical precision function PP is not concave. In particular, we mention now that both Theorem 5 and Proposition 6 do not necessarily hold in the empirical setting. Still, as suggested, we conducted an empirical search for the maximum gain from under-reporting, given the fitted precision functions PP. The revised article contains a new section, subsection 5.4, dedicated to this new experiment. For several sets of parameters and different scenarios of strategic firms, all experiments suggest that no large gain on utility can be obtained by under-reporting when an integral version of our contract HH is implemented. Indeed, Table 1 in the revised article presents the maximum gain that any firm can obtain by lying in the different scenarios. We find that when one firm is strategic while the rest are truthful, the maximum gain by under-reporting corresponds to 0.03800.0380%. Similarly, when two firms are strategic while the rest are truthful, the maximum gain does not surpass 1.41631.4163%.

Even though the empirical results suggest a close-to-strategyproofness of the contract HH, we recognize an important limitation in our article from the need of precision functions being concave. We thank the reviewer(s) for remarking this limitation, which we have added to the limitations section of the paper.

Comments on the paper

1 comment

Dear Authors and Reviewers,

Thank you for supporting ICLR'27!

We are now entering the discussion phase. Please make the most of this time to exchange your views. I encourage each Reviewer, if you have not already done so, to read the other reviews and the Authors' responses. Please respond to the Authors and actively participate in the discussion to help clarify any remaining issues.

Thank you,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.