Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Deterministic-Allocation and Anonymous Joint Advertising in E-commerce Platforms

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read JTransNet, a transformer-based neural auction mechanism, generates anonymous and deterministic allocations for joint e-commerce advertising, and the paper reports it beats VCG, JAMA, and prior AMD baselines on revenue while staying…

desk verdict A genuinely deployed deterministic/anonymous AMD architecture with a real A/B lift, undermined by an overstated theorem and regret measured only on a truncated misreport grid. read the letter →

arxiv 2506.02435 v4 pith:6QWEUX2J submitted 2025-06-03 cs.GT

Zhen Zhang , Luowen Liu , Wanzhi Zhang , Zitian Guo , Kun Huang , Qi Qi , Qianlong Xie , Xingxing Wang This is my paper · ORCID
classification cs.GT MSC 91B2691B0368T07
keywords automatedmechanismdesigndeterministicallocationanonymityjointauctione-commerceadvertisingtransformernear-DSICneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that earlier automated mechanism design (AMD) networks trained to produce near-truthful optimal auctions output probabilistic allocations, and in online advertising this is usually infeasible when every ad slot must be filled: Theorem 1 shows the probability matrices those networks generate almost never correspond to any distribution over valid slot assignments. To fix this, the paper introduces JTransNet, a transformer-based network that produces deterministic 0-1 allocations by sorting bundle scores, and that guarantees anonymity by using a transformer without positional encoding, so the outcome depends only on bid values, not on who bids or in what order. Payments are set below each bidder's expected gain, which keeps the mechanism individually rational by construction. On simulated data, on real auction logs with 10 brands, 10 stores, and 5 slots, and in a one-month live A/B test on 40% of traffic, JTransNet achieves significantly higher platform revenue than VCG, JAMA, and prior AMD baselines with ex-post regret below 0.130 on real data. The authors claim this is the first AMD architecture for near-DSIC optimal mechanisms deployed in real industrial scenarios.

What carries the argument

The load-bearing object is the bundle allocation matrix S built from a descending sort of bundle scores. Each bundle's score is the sum of its brand bid and store bid; the k-th largest score wins slot k, producing a permutation-like 0-1 matrix that enforces 'one bundle per slot' and 'one slot per bundle' by construction. During training the argmax in the sort is replaced with a softmax operator at temperature τ so gradients can flow, and at test time the exact argmax is applied, yielding a genuinely deterministic allocation. Payments are derived by multiplying a sigmoid gate in [0,1] by the bidder's value times CTR times the obtained allocation, which keeps payment below the bidder's expected gain and enforces IR; the training objective is an augmented Lagrangian of negative revenue plus per-bidder ex-post regret, with the best misreport found by enumerating a grid Γ = {0, 0.05, ..., 1.45} of multiplicative coefficients.

What would settle it

Evaluate JTransNet with a continuous misreport search (for example, fine-grained grid or gradient ascent over [0, v_max] for each bidder) and check whether per-bidder ex-post regret stays near zero; if regret rises substantially beyond the reported values, the near-DSIC claim fails. A separate check for Theorem 1: take a RegretNet-style probabilistic allocation for 3 bundles and 2 slots whose row and column sums are all ≤ 1 but with each column sum strictly below 1, and verify that no convex combination of the 6 feasible deterministic assignments reproduces it, demonstrating the promised infeasibility gap.

Watch

Extended reading notes

Core claim

JTransNet's central move is to build the allocation rule as a hard 0-1 assignment rather than a probability matrix. Bundle c receives slot k if it has the k-th largest score among all bundles, where each bundle's score is computed from the bids of its constituent brand and store; the result is a matrix with at most one 1 per row and column, automatically satisfying the two feasibility constraints of a joint ad auction. Because sorting with argmax is non-differentiable, training uses a softmax relaxation with temperature τ and the exact argmax at test time. The paper proves that this determinism is necessary: Theorem 1 shows that the row-sum ≤ 1 and column-sum ≤ 1 constraints used by prior RegretNet-style architectures cannot in general be reconciled with the requirement that all K slots be allocated, so their probabilistic outputs admit no feasible deterministic rounding. Anonymity is obtained by processing bids through a transformer without positional encoding, and payments are computed as a sigmoid-scaled fraction of each bidder's expected value for the slot actually won, keeping payment strictly below value so individual rationality holds automatically. The empirical claim is that on simulated and real data JTransNet's revenue exceeds VCG, JAMA, and RegretNet-based baselines with only a small loss in social welfare relative to VCG, and that the online A/B test gave a +1.87% revenue lift.

Load-bearing premise

The near-truthfulness claim depends on the finite misreport list {0, 0.05, 0.1, ..., 1.45} used during evaluation containing every bidder's best possible cheat; if the true optimal misreport is outside that list or has a different functional form, the reported regret understates the true regret and the mechanism may not be near-truthful.

Editorial extensions

If this is right

  • In online advertising, any AMD architecture that encodes allocation as a probability matrix with only row and column sum constraints will fail for the vast majority of instances when all slots must be allocated; JTransNet's sorting-based 0-1 module removes that obstruction and can be ported to other non-deterministic AMD architectures with minor modifications.
  • JTransNet achieves higher platform revenue than VCG, JAMA, and prior AMD baselines in simulated settings and on real auction logs (10 brands, 10 stores, 5 positions), while keeping average regret low and social welfare close to the VCG benchmark.
  • In a one-month online A/B test on 40% of an e-commerce platform's traffic, JTransNet increased revenue by +1.87% relative to the previously used joint auction mechanism, with average inference time under 10 ms and TP99 at 15 ms, and is now fully deployed.
  • Anonymity—the property that outcomes depend only on bid values, not on bidder identities or bid order—can be achieved in a joint auction by a transformer without positional encoding, without sacrificing deterministic allocation or near-truthfulness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer that the infeasibility result applies beyond joint auctions to any probabilistic AMD mechanism for sponsored search or display ads, because the same per-slot and per-advertiser constraints hold; the sorting-based deterministic construction may serve as a general recipe for making AMD mechanisms deployable.
  • A testable extension is to anneal the temperature τ during training and measure whether the final exact-argmax allocation's regret and revenue remain stable; this would quantify how much the softmax approximation used for gradients biases the evaluated mechanism.
  • We infer that the revenue advantage over exact-IC JAMA reflects a deliberate trade-off of strict truthfulness for higher revenue, which is reasonable in practice when advertisers face search costs in finding optimal misreports; a follow-up could test whether an exact-IC deterministic variant loses the reported revenue gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes JTransNet, a transformer-based automated mechanism design (AMD) architecture for joint advertising auctions, which aims to satisfy near-DSIC and IR while producing deterministic, anonymous allocations. The authors prove, in a theorem, that non-deterministic AMD outputs are often infeasible when all ad slots must be filled, and they introduce a deterministic top-K allocation module together with a payment network. Offline experiments compare JTransNet with RegretNet, VCG, and JAMA on simulated and real auction data, and an online A/B test reports a +1.87% revenue lift over the platform's previous joint auction mechanism. The central evidence for near-DSIC is an empirical regret computed over the finite misreport grid Γ = {0, 0.05, ..., 1.45}.

Significance. If the near-DSIC claim can be substantiated, JTransNet would be a meaningful practical contribution: it is a deployed AMD mechanism with deterministic allocation and anonymity, and it reports substantial revenue gains over VCG and JAMA while preserving most social welfare. The deterministic top-K module and the anonymity-by-transformer design are reusable ideas, and the online deployment details (inference latency, TP99, one-month A/B test) are concrete evidence of industrial viability. However, the theoretical infeasibility theorem is currently overstated, and the near-DSIC evaluation is a lower bound rather than an upper bound on regret because of the truncated misreport grid. The value of the paper therefore depends on whether the incentive guarantees can be re-established with a sound regret evaluation.

major comments (3)
  1. [§3.2, §4 'Hyperparameters', and Eq. (2)] The reported regret is computed by enumerating μ ∈ Γ = {0, 0.05, ..., 1.45} and using the misreport min(μv_i, v_i^max). For any true value v_i < v_i^max / 1.45, all misreports in the interval (1.45v_i, v_i^max] are never tested. Under the simulated distributions U[0,1], this affects about 31% of types, and for v_i = 0.5 the feasible misreport 1.0 is simply omitted. Since the test-time allocation is a hard top-K rule, the utility of a bidder as a function of the misreport is piecewise constant with jump discontinuities at sorting thresholds, so the maximizing misreport can lie at one of the untested values. Consequently, the regret figures in Tables 1 and 2 are lower bounds on the true ex-post regret, not upper bounds, and the paper provides no Lipschitz, monotonicity, or grid-refinement argument to bound the gap. The near-DSIC claim is load-bearing, and it must be supported either by a substantially larger and unbounded misreport grid, by a randomized or exhaustive search over the type space, or by a theoretical argument that the optimal misreport is contained in the tested family. In addition, Algorithm 1 lines 10 and 13 set the misreport to r v_i without the min(·, v_max) truncation used in the text, which would create infeasible bids when r v_i exceeds the type bound.
  2. [§2.1, Theorem 1, and Appendix A] The proof in Appendix A shows that if every ad slot is allocated with probability 1, then the column sums of the allocation matrix must each equal 1, whereas the constraint set considered only enforces column sums ≤ 1. This is correct but definitional; it does not establish the abstract's claim that previous non-deterministic AMD methods 'lead to the non-existence of feasible solutions in the vast majority of cases.' The argument relies on the unstated and unsupported assumption that mainstream AMD architectures generate exactly the set of matrices with row and column sums at most 1. Many practical AMD allocations are produced by per-slot softmax with a dummy option, and if the dummy probability is trained to zero on every slot, the resulting matrix has column sums equal to 1 and is feasible. The theorem should be restated as a precise condition on column sums, and the assertion about 'mainstream AMD architectures' should either be proven, cited, or omitted.
  3. [§4.2, §4.3, and §5] The empirical support for near-DSIC is incomplete even apart from the grid issue. In Table 2, the regret is reported only as '<0.130' without exact values, standard errors, or a comparison of regret across the three real-data settings, so the reader cannot assess whether the mechanism is meaningfully close to DSIC. The online A/B test in §4.3 measures revenue relative to the platform's previous mechanism, not incentive compatibility, and the discussion that advertisers cannot find the optimal misreport is a behavioral observation that does not bound worst-case regret. Since near-DSIC is a central claimed contribution, the paper should report exact regret values on a sufficiently rich misreport search and state explicitly that the online revenue lift does not measure incentive compatibility.
minor comments (5)
  1. [§2.2, Eq. (2)] The notation rgti·(v) conditions on the full value profile v, but the empirical definition then averages over samples from F; the formal definition should make explicit that the expectation is over the other bidders' values conditional on v_i, or the notation should be changed accordingly.
  2. [Table 1 and §4 'Baseline Methods'] The text says RegretNet is used only for the traditional ad setting, yet Table 1 reports RegretNet in the joint-ad settings A–D; the role of RegretNet in these comparisons should be clarified.
  3. [§4.3 'Online A/B Test'] The online A/B test would be more convincing with details on the control mechanism, the number of auctions or impressions, the length of the pre-period, and confidence intervals for the +1.87% lift.
  4. [Table 3] The sensitivity analysis varies τ and the initial ρ but reports only revenue; reporting regret under the same variations would be more informative, especially because the train-time soft top-K relaxation and the test-time hard top-K rule may behave differently.
  5. [Abstract and Introduction] The phrases 'vast majority of cases' and 'non-existence of feasible solutions' are not quantified in the theorem or its proof; the language should be aligned with the actual mathematical statement, which is a measure-zero or constraint-satisfaction statement.

Circularity Check

1 steps flagged · score 6.0 of 10

The near-DSIC claim is evaluated on the same finite misreport grid used to train the network, so the reported regret is the training objective rather than an independent upper bound; revenue gains vs VCG/JAMA remain externally grounded.

  1. fitted input called prediction [Section 3.2 (Computing the optimal misreports; Algorithm 1, lines 7-15) and Section 4 (Hyperparameters)]
    "The regret of JTransNet is calculated by enumeration [33]. In this enumeration method, we calculate the misreport using a coefficient µ∈Γ ... the coefficient µ∈Γ that maximizes the utility of that brand or store is used to compute the misreport and regret ... In addition, for the test of JTransNet, we use the coefficient set Γ = {0,0.05,0.1,0.15,...,1.45} to obtain the optimal misreports for calculating the regret."

    Algorithm 1 trains JTransNet by enumerating r∈Γ and selecting the best misreport in Γ to compute regret, then Section 4 reports regret using the same Γ. The reported near-DSIC quantity is therefore the value of the training surrogate, not an upper bound over the full value space. Because Γ is capped at 1.45, for any v < vmax/1.45 the deviations in (1.45v, vmax] are never searched; hard top-K allocation has upward jumps at sorting thresholds, so the untested region can contain strictly larger utility gains. The paper supplies no Lipschitz, monotonicity, or grid-refinement argument bounding the gap. Thus the central near-DSIC claim reduces to the grid on which the network was optimized.

full rationale

The only significant circularity concern is the DSIC evaluation: the regret numbers in Tables 1-2 are computed over the same finite coefficient set Γ used in training, so they are a lower-bound surrogate rather than an independent verification of near-DSIC. I do not find circularity in Theorem 1, which is a valid (if narrow) observation about probability matrices with row/column sums bounded by one not representing full allocation; nor in the revenue comparisons against VCG and JAMA, which use external mechanisms and held-out data; nor in the online A/B revenue lift, which measures deployment revenue and not DSIC. The self-citations to prior joint-auction work are background and do not carry the derivation. However, because the paper's flagship incentive-compatibility claim is evaluated with the same grid that defines the training objective, and no argument shows the grid contains optimal misreports, the near-DSIC claim is partially forced by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central empirical claim (JTransNet improves revenue) rests on the parameterized mechanism, which is a standard AMD setup. The distinctive theoretical claim (Theorem 1) rests on a modeling assumption about prior architectures and a trivial consequence of the all-slots-filled constraint.

free parameters (4)
  • network weights w = learned
    All mechanism parameters (allocation scoring, payment scaling) are optimized on training data to maximize revenue subject to regret penalties.
  • softmax temperature tau = 0.001 (with robustness checks)
    Controls the approximation error of the differentiable sorting step; chosen by hand and tested across 0.0009 to 0.0011 in Table 3.
  • penalty factor rho = 0.603 (initial)
    Augmented Lagrangian penalty; initial value chosen by hand and tested in Table 3.
  • misreport coefficient grid Gamma = {0, 0.05, ..., 1.45}
    Used to compute optimal misreports and regret; influences the learned payments and the reported regret.
assumptions (6)
  • standard math Differentiable sorting approximation converges to exact sorted order as tau approaches 0
    Invoked in Section 3.1 to make the deterministic allocation differentiable during training; relies on Grover et al. [32].
  • domain assumption Advertiser values are drawn from a known distribution F and bidders maximize expected utility
    Invoked in Section 2.1 and used in the empirical regret estimation (Equation 2).
  • domain assumption Click-through rates alpha_k are known, exogenous, and decreasing in slot rank
    Invoked in Section 2.1; allocation values and payments are linear in alpha_k.
  • domain assumption All ad slots must be allocated in every auction
    This is the premise of Theorem 1 and the motivation for deterministic allocation; stated in the abstract and Section 2.
  • ad hoc to paper Mainstream AMD architectures generate exactly the set of allocation matrices with row and column sums at most 1
    Assumed in Appendix A for Theorem 1; this characterization of prior AMD outputs is asserted, not derived.
  • ad hoc to paper The misreport grid Gamma is sufficient to find optimal misreports
    Assumed in Section 3.2 and used in both training and evaluation; not justified by analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deterministic-Allocation and Anonymous Joint Advertising in E-commerce Platforms." pith.science (2026). https://pith.science/paper/6QWEUX2J

@misc{pith2026250602435,
  author       = {Pith},
  title        = {Pith review of: Deterministic-Allocation and Anonymous Joint Advertising in E-commerce Platforms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6QWEUX2J}},
  note         = {Machine review of arXiv:2506.02435}
}
read the original abstract

With the advancement of machine learning, an increasing number of studies are employing automated mechanism design (AMD) methods for optimal auction design. However, all previous AMD architectures designed to generate optimal mechanisms that satisfy near dominant strategy incentive compatibility (DSIC) fail to achieve deterministic allocation, and some also lack anonymity, thereby impacting the efficiency and fairness of advertising allocation. This has resulted in a notable discrepancy between the previous AMD architectures for generating near-DSIC optimal mechanisms and the demands of real-world advertising scenarios. In this paper, we prove that in all online advertising scenarios, when all ad slots must be allocated, previous non-deterministic allocation AMD methods lead to the non-existence of feasible solutions in the vast majority of cases, resulting in a gap between the rounded solution and the optimal solution. Furthermore, we propose JTransNet, a transformer-based neural network architecture, designed for optimal deterministic-allocation and anonymous joint auction design. Although the deterministic allocation module in JTransNet is designed for the latest joint auction scenarios, it can be applied to other non-deterministic AMD architectures with minor modifications. Additionally, our offline and online data experiments demonstrate that, in joint auction scenarios, JTransNet significantly outperforms the considered baselines in terms of platform revenue.

Figures

Figures reproduced from arXiv: 2506.02435 by the authors.

Figure 1
Figure 1. Comparison between Traditional Ad Model and Joint Ad Model. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The JTransNet architecture for the setting of [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 2
Figure 2. Computing the bundles’ allocation matrix S. There are several constraints that need to be implemented through the bundles’ allocation matrix S. Therefore, JTransNet begins by computing the bundles’ allocation matrix S, followed by determining the bidders’ allocation matrix A based on S. The matrix S contains the allocation probabilities for each bundle across different ad positions. Specifically, the allocation of b… view at source ↗
Figures from the paper (2 more)
Figure 2
Figure 2. Figure 2: Ci· and C·j represent the sets of all bundles containing brand i and store j, respectively. The total payments of brand i and store j can be expressed as pi· = P c∈Ci· pi(c)· and p·j = P c∈C·j p·j(c) , respectively. When DSIC is ensured, since the utility of the brand …
Figure 3
Figure 3. Figure 3: The bundle relationships between store bidders and brand bidders for Settings [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 25 canonical work pages

  1. [1]

    Position auctions.international Journal of industrial Organization, 25(6):1163–1178, 2007

    Hal R Varian. Position auctions.international Journal of industrial Organization, 25(6):1163–1178, 2007

    work page 2007

  2. [2]

    Online ad auctions.American Economic Review, 99(2):430–434, 2009

    Hal R Varian. Online ad auctions.American Economic Review, 99(2):430–434, 2009

    work page 2009

  3. [3]

    Joint auction in the online advertising market

    Zhen Zhang, Weian Li, Yahui Lei, Bingzhe Wang, Zhicheng Zhang, Qi Qi, Qiang Liu, and Xingxing Wang. Joint auction in the online advertising market. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4362–4373, 2024

    work page 2024

  4. [4]

    Optimal auctions through deep learning

    Paul Dütting, Zhe Feng, Harikrishna Narasimhan, David Parkes, and Sai Srivatsa Ravindranath. Optimal auctions through deep learning. InInternational Conference on Machine Learning, pages 1706–1715. PMLR, 2019

    work page 2019

  5. [5]

    A context- integrated transformer-based neural network for auction design

    Zhijian Duan, Jingwu Tang, Yutong Yin, Zhe Feng, Xiang Yan, Manzil Zaheer, and Xiaotie Deng. A context- integrated transformer-based neural network for auction design. InInternational Conference on Machine Learning, pages 5609–5626. PMLR, 2022

    2022

  6. [6]

    Deep learning for revenue-optimal auctions with budgets

    Zhe Feng, Harikrishna Narasimhan, and David C Parkes. Deep learning for revenue-optimal auctions with budgets. InProceedings of the 17th International Conference on Autonomous Agents and Multiagent Systems, pages 354–362, 2018

    work page 2018

  7. [7]

    A permutation-equivariant neural network architecture for auction design

    Jad Rahme, Samy Jelassi, Joan Bruna, and S Matthew Weinberg. A permutation-equivariant neural network architecture for auction design. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 5664–5672, 2021

    work page 2021

  8. [8]

    Internet advertising and the generalized second- price auction: Selling billions of dollars worth of keywords.American economic review, 97(1):242–259, 2007

    Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. Internet advertising and the generalized second- price auction: Selling billions of dollars worth of keywords.American economic review, 97(1):242–259, 2007

    work page 2007

Show all 39 references
  1. [9]

    On revenue in the generalized second price auction

    Brendan Lucier, Renato Paes Leme, and Éva Tardos. On revenue in the generalized second price auction. In Proceedings of the 21st international conference on World Wide Web, pages 361–370, 2012

    2012

  2. [10]

    On the efficiency of equilibria in generalized second price auctions

    Ioannis Caragiannis, Christos Kaklamanis, Panagiotis Kanellopoulos, and Maria Kyropoulou. On the efficiency of equilibria in generalized second price auctions. InProceedings of the 12th ACM conference on Electronic commerce, pages 81–90, 2011

    2011

  3. [11]

    Bayes-nash equilibria of the generalized second price auction

    Renato D Gomes and Kane S Sweeney. Bayes-nash equilibria of the generalized second price auction. In Proceedings of the 10th ACM conference on Electronic Commerce, pages 107–108, 2009

    2009

  4. [12]

    Revenue analysis of a family of ranking rules for keyword auctions

    Sébastien Lahaie and David M Pennock. Revenue analysis of a family of ranking rules for keyword auctions. In Proceedings of the 8th ACM Conference on Electronic Commerce, pages 50–56, 2007

    2007

  5. [13]

    Revenue optimization in the generalized second-price auction

    David RM Thompson and Kevin Leyton-Brown. Revenue optimization in the generalized second-price auction. InProceedings of the fourteenth ACM conference on Electronic commerce, pages 837–852, 2013

    2013

  6. [14]

    Optimal auction design.Mathematics of operations research, 6(1):58–73, 1981

    Roger B Myerson. Optimal auction design.Mathematics of operations research, 6(1):58–73, 1981

    1981

  7. [15]

    Counterspeculation, auctions, and competitive sealed tenders.The Journal of finance, 16(1):8–37, 1961

    William Vickrey. Counterspeculation, auctions, and competitive sealed tenders.The Journal of finance, 16(1):8–37, 1961

    1961

  8. [16]

    Multipart pricing of public goods.Public choice, pages 17–33, 1971

    Edward H Clarke. Multipart pricing of public goods.Public choice, pages 17–33, 1971

    1971

  9. [17]

    Incentives in teams.Econometrica: Journal of the Econometric Society, pages 617–631, 1973

    Theodore Groves. Incentives in teams.Econometrica: Journal of the Econometric Society, pages 617–631, 1973

    1973

  10. [18]

    Mixed bundling auctions.Journal of Economic Theory, 134(1):494–512, 2007

    Philippe Jehiel, Moritz Meyer-Ter-Vehn, and Benny Moldovanu. Mixed bundling auctions.Journal of Economic Theory, 134(1):494–512, 2007

    2007

  11. [19]

    Mixed-bundling auctions with reserve prices

    Pingzhong Tang and Tuomas Sandholm. Mixed-bundling auctions with reserve prices. InAAMAS, pages 729–736, 2012

    2012

  12. [20]

    Complexity of mechanism design

    Vincent Conitzer and Tuomas Sandholm. Complexity of mechanism design. InProceedings of the Eighteenth conference on Uncertainty in artificial intelligence, pages 103–110, 2002

    2002

  13. [21]

    Self-interested automated mechanism design and implications for optimal combinatorial auctions

    Vincent Conitzer and Tuomas Sandholm. Self-interested automated mechanism design and implications for optimal combinatorial auctions. InProceedings of the 5th ACM Conference on Electronic Commerce, pages 132–141, 2004

    2004

  14. [22]

    Automated design of revenue-maximizing combinatorial auctions

    Tuomas Sandholm and Anton Likhodedov. Automated design of revenue-maximizing combinatorial auctions. Operations Research, 63(5):1000–1025, 2015

    2015

  15. [23]

    Differentiable economics for randomized affine maximizer auctions.arXiv preprint arXiv:2202.02872, 2022

    Michael Curry, Tuomas Sandholm, and John Dickerson. Differentiable economics for randomized affine maximizer auctions.arXiv preprint arXiv:2202.02872, 2022

    2022 arXiv

  16. [24]

    A scalable neural network for dsic affine maximizer auction design.Advances in Neural Information Processing Systems, 36:56169–56185, 2023

    Zhijian Duan, Haoran Sun, Yurong Chen, and Xiaotie Deng. A scalable neural network for dsic affine maximizer auction design.Advances in Neural Information Processing Systems, 36:56169–56185, 2023. 13

    2023

  17. [25]

    Automated mechanism design via neural networks.arXiv preprint arXiv:1805.03382, 2018

    Weiran Shen, Pingzhong Tang, and Song Zuo. Automated mechanism design via neural networks.arXiv preprint arXiv:1805.03382, 2018

    2018 arXiv

  18. [26]

    Gemnet: Menu-based, strategy-proof multi-bidder auctions through deep learning.arXiv preprint arXiv:2406.07428, 2024

    Tonghan Wang, Yanchen Jiang, and David C Parkes. Gemnet: Menu-based, strategy-proof multi-bidder auctions through deep learning.arXiv preprint arXiv:2406.07428, 2024

    2024 arXiv

  19. [27]

    Preferencenet: Encoding human preferences in auction design with deep learning.Advances in Neural Information Processing Systems, 34:17532–17542, 2021

    Neehar Peri, Michael Curry, Samuel Dooley, and John Dickerson. Preferencenet: Encoding human preferences in auction design with deep learning.Advances in Neural Information Processing Systems, 34:17532–17542, 2021

    2021

  20. [28]

    Optimal-er auctions through attention

    Dmitry Ivanov, Iskander Safiulin, Igor Filippov, and Ksenia Balabaeva. Optimal-er auctions through attention. Advances in Neural Information Processing Systems, 35:34734–34747, 2022

    2022

  21. [29]

    Selling joint ads: A regret minimization perspective.arXiv preprint arXiv:2409.07819, 2024

    Gagan Aggarwal, Ashwinkumar Badanidiyuru, Paul Dütting, and Federico Fusco. Selling joint ads: A regret minimization perspective.arXiv preprint arXiv:2409.07819, 2024

    2024 arXiv

  22. [30]

    Hybrid advertising in the sponsored search

    Zhen Zhang, Weian Li, Yuhan Wang, Qi Qi, and Kun Huang. Hybrid advertising in the sponsored search. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2463–2472, 2025

    2025

  23. [31]

    Joint bidding in ad auctions

    Yuchao Ma, Weian Li, Wanzhi Zhang, Yahui Lei, Zhicheng Zhang, Qi Qi, Qiang Liu, and Xingxing Wang. Joint bidding in ad auctions. InAnnual Conference on Theory and Applications of Models of Computation, pages 344–354. Springer, 2024

    2024

  24. [32]

    Stochastic optimization of sorting networks via continuous relaxations.arXiv preprint arXiv:1903.08850, 2019

    Aditya Grover, Eric Wang, Aaron Zweig, and Stefano Ermon. Stochastic optimization of sorting networks via continuous relaxations.arXiv preprint arXiv:1903.08850, 2019

    1903 arXiv

  25. [33]

    Nma: Neural multi-slot auctions with externalities for online advertising.arXiv preprint arXiv:2205.10018, 2022

    Guogang Liao, Xuejian Li, Ze Wang, Fan Yang, Muzhi Guan, Bingqi Zhu, Yongkang Wang, Xingxing Wang, and Dong Wang. Nma: Neural multi-slot auctions with externalities for online advertising.arXiv preprint arXiv:2205.10018, 2022. 14 A Proof of Theorem 1 A.1 An Example First, cons...

    2022 arXiv

  26. [34]

    The 1st bundle is assigned to the 1st slot, and the 2nd bundle is assigned to the 2nd slot

  27. [35]

    The 1st bundle is assigned to the 1st slot, and the 3rd bundle is assigned to the 2nd slot

  28. [36]

    The 1st bundle is assigned to the 2nd slot, and the 2nd bundle is assigned to the 1st slot

  29. [37]

    The 1st bundle is assigned to the 2nd slot, and the 3rd bundle is assigned to the 1st slot

  30. [38]

    The 2nd bundle is assigned to the 1st slot, and the 3rd bundle is assigned to the 2nd slot

  31. [39]

    Let P r1, P r2,

    The 2nd bundle is assigned to the 2nd slot, and the 3rd bundle is assigned to the 1st slot. Let P r1, P r2, . . . , P r6 represent the probabilities of scenarios 1, 2, ..., 6 occurring. To ensure that all probabilities in the allocation probability matrixS ′ are satisfied, the...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.