Pith. sign in

REVIEW 4 major objections 5 minor 31 references

A Reliable Vertical Federated Learning Framework for Traffic State Estimation with Data Selection and Incentive Mechanisms

T0 review · 4 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 A vertical federated learning framework for traffic state estimation can be made reliable in both training and operation by using a neural mutual-information model to select the most informative data providers and a penalty-based…

desk verdict The MI-based provider selection is a solid, near-Oracle contribution; the incentive mechanism's convergence proof assumes its own conclusion, and the headline accuracy numbers don't match Table V. read the letter →

arxiv 2506.01285 v1 pith:624Y3KER submitted 2025-06-02 cs.GT

Zijun Zhan , Yaxian Dong , Daniel Mawunyo Doe , Yuqing Hu , Shuai Li , Shaohua Cao , Zhu Han This is my paper
classification cs.GT
keywords verticalfederatedlearningtrafficstateestimationmutualinformationdataproviderselectionincentivemechanismdesignsupervisiongamepenalty-based
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 sets out to make vertical federated learning (VFL) for traffic state estimation—predicting traffic flow and density from data split across a municipal authority and several mobility providers—dependable in both phases of the model's life. Its first claim is that a small neural mutual-information model trained on drone-imaged labels can score how informative each provider's data is for each road segment, letting the authority pick the single best provider per segment without seeing raw data. Its second claim is that the same mutual-information model can inspect providers during operation, enabling a supervision game with escalating penalties that drives both the providers' sloth probability and the authority's inspection probability to zero. If both claims hold, VFL-based traffic forecasting can select competent data partners and deter lazy behavior, which the simulations support: about 11% better density prediction, about 21% better flow prediction, and $130–400 higher authority utility than the benchmarks.

What carries the argument

The load-bearing object is the neural mutual-information model $I_{\theta_{MI}}(X;Y)$, trained through the Donsker–Varadhan variational lower bound so that it approximates the KL divergence between the joint distribution of provider data and labels and the product of their marginals. This model outputs the quality score $Q(X_{k,n})\approx I^n(X_{k,n};Y)$ used to solve the provider-selection problem $\max_A \sum_{k,n} a_{k,n} Q(X_{k,n})$ subject to one provider per road segment. The second half of the machinery is the penalty-based supervision game: a repeated game between the authority and each provider governed by the recurrences $1/\gamma_{i+1} = 1/\gamma_i + \gamma_i\epsilon_i\rho(\beta-1)/S - W/S$ and $1/\epsilon_{i+1} = 1/\epsilon_i + \gamma_i\epsilon_i\rho(\beta-1)/(H-H') - W/(H-H')$, where $\gamma$ is the provider's laziness probability, $\epsilon$ is the authority's inspection probability, $\rho$ is the base penalty, $\beta$ the escalation multiplier for repeat offenders, $W$ the reward, $S$ the inspection cost, and $H-H'$ the provider's saved cost when lazy. These recurrences produce the theoretical upper bounds in Propositions 1 and 2.

What would settle it

Simulate the recurrences (38) and (43) over a dense grid of feasible parameters ($\gamma_0 \in [0,1]$, $\epsilon_0 \in [0,1]$, $\rho$, $\beta$, $W$, $S$) and check whether every trajectory hits one of the terminal conditions in Eq. (44) at some finite cycle. A single trajectory that cycles or plateaus without hitting any terminal condition—while still converging to nonzero probabilities—would show that Proposition 1's upper bound is not established for that regime, and if in that trajectory the observed $\gamma$ or $\epsilon$ exceeds the bound, the mechanism's guarantee is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that one trained artifact—a neural estimator of mutual information between a provider's traffic data and the authority's label—can carry the entire reliability story of a VFL traffic-state-estimation system. Before training, the estimator ranks providers per road segment, so the authority selects the most competent one (addressing Q1). After training, the same estimator acts as an inspection oracle that detects whether a provider is supplying real-time data or lazy substitutes such as randomly generated or historical data; with that oracle in place, the authority and providers play a repeated supervision game in which a provider caught being lazy on two consecutive settlement cycles pays an additional penalty of $(\beta-1)\rho$. The paper proves upper bounds on the long-run sloth probability $\gamma$ and inspection probability $\epsilon$ (Propositions 1 and 2) and reports convergence of both to zero under a wide range of settings, which it presents as guaranteeing the authority's utility and resolving Q2. The numerical claims attached to this discovery are an 11.23% gain in traffic-density accuracy, a 21.15% gain in traffic-flow accuracy over the random benchmark, and a $130–400 increase in the authority's utility over the collective-learning baseline.

Load-bearing premise

The convergence guarantee of the incentive mechanism assumes that after finitely many settlement cycles the repeated game reaches one of three terminal conditions—$\gamma_i\epsilon_i\rho(\beta-1)=W$, $\gamma_i=0$, or $\epsilon_i=0$—and the paper does not prove that any of these conditions is actually reached.

Editorial extensions

If this is right

  • VFL-based traffic state estimation becomes reliable end-to-end: competent providers are chosen before training, and lazy data provision is deterred during operation by the same mutual-information model.
  • The two research questions Q1 and Q2 are answered by a single artifact, so the marginal cost of policing after training is nearly zero once the selection model exists.
  • Because the additional penalty applies only to providers caught lazy twice consecutively, the mechanism avoids the two failure modes of a large static penalty: the authority's budget is not breached and honest providers are not deterred from participating.
  • In the simulated pNEUMA experiments the framework's inspection scheme detects every tested lazy strategy—random and historical substitutions at 20–100% intensity—as negative mutual information, so the authority can spot sloth regardless of the provider's evasion tactic.

Reading between the lines

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

  • The same two-role mutual-information design could generalize to other vertical federated learning settings where a label owner aggregates features from multiple providers: the estimator is a data-quality gate, and the supervision game is a generic deterrent, independent of traffic-specific structure.
  • The theoretical convergence of the incentive mechanism rests on the unproven assumption that the recurrence actually hits one of the terminal conditions in Eq. (44); a natural follow-up is to characterize the parameter regimes (initial $\gamma_0$, $\epsilon_0$, $\rho$, $\beta$, $W$, $S$) under which the game provably reaches those conditions, and to test whether the claimed upper bound holds outsi
  • The selection stage assumes the MI model trained on drone-imaged labels faithfully captures the value of each provider's data; if the MI estimate is noisy or the label distribution shifts over time, selection quality could degrade, suggesting adaptive re-scoring as a testable extension.
  • The utility model ignores the negligible cost of uploading intermediate results and treats each provider's contribution as equal ($1/K$); in settings with heterogeneous segment importance, weighting selection and inspection by segment value would be a direct extension.
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

4 major / 5 minor

Summary. The paper proposes a two-part framework for vertical federated learning (VFL) traffic state estimation: (i) mutual-information (MI) based data provider selection among multiple mobility providers (MPs) per road segment, and (ii) a penalty-based supervision game incentive mechanism intended to drive the lazy-provider probability and the MA inspection probability to zero. The MI model is trained by the municipal authority (MA) with drone imaging data, split across entities, and used both for provider selection and for detecting low-quality data during inference. The incentive part models MA-MP interaction as a repeated game with an additional penalty for repeated sloth, deriving equilibrium probabilities and theoretical upper bounds. Experiments on the pNEUMA dataset compare selection schemes (Central, Random, Oracle, VFLFS) and incentive benchmarks (BCL, SGF), reporting improved prediction accuracy and MA utility gains.

Significance. If the theoretical claims were fully supported, the work would be a useful integration of feature-selection-style participant screening with a supervision-game incentive mechanism for VFL traffic estimation. The MI selection evaluation is the strongest part: Table IV shows monotonic decrease of MI with injected noise, and Table V shows near-Oracle performance on density and flow MAE/RMSE. The idea of reusing the trained MI model for both selection and inspection is coherent and the empirical selection study is convincing. However, the headline guarantee of the incentive mechanism currently rests on an unproven and partially circular argument, so the central claim cannot be accepted as it stands.

major comments (4)
  1. [Section V-B, Eqs. (44)-(52)] The convergence proof assumes, rather than establishes, that one of the terminal conditions in (44) is reached at some round i*. Even granting that assumption, the key inequality (50) — γ_{i*-1} ε_{i*-1} ρ(β-1) > W with γ_{i*-1} > 0 and ε_{i*-1} > 0 — is asserted without derivation. The recurrence (38) shows that 1/γ_{i+1} increases only when the bracket is positive; when γ_i ε_i ρ(β-1) < W, γ can increase. Thus Propositions 1 and 2 do not establish the claimed convergence to zero, and the statement that the mechanism 'guarantees the utility of MA' is unsupported.
  2. [Section VI-C3, Fig. 9] The 'theoretical upper bound' is computed by imposing γ_i* ε_i* ρ(β-1) = W, which is exactly the equality the proof is supposed to derive from (50); moreover (50) states a strict inequality, not an equality. The comparison between the simulated probability gap and this imposed bound is therefore circular and cannot serve as empirical validation of Propositions 1 and 2.
  3. [Section VI-B, Table V vs abstract/conclusion] The accuracy-improvement numbers are internally inconsistent. Table V yields density MAE improvement (29.465−26.146)/29.465 = 11.26% and flow MAE improvement (21.100−17.498)/21.100 = 17.07%, while the abstract reports 11.23% and 23.15%, and the conclusion reports 11.23% and 21.15%. The paper should state one consistent set of numbers and recompute them from Table V.
  4. [Section V-A, after Eq. (26)] The assumption π' = 0 (a lazy MP yields zero profit for the MA) is introduced without justification and is used directly to obtain the closed-form equilibria (29) and (30). Since the parameter regime and the utility advantages in Figs. 12-13 depend on these formulas, the sensitivity of the conclusions to relaxing this assumption should be analyzed, or the assumption should be justified from the data.
minor comments (5)
  1. [Abstract and Section VII] The abstract states the framework augments 'traffic flow and density prediction accuracy by 11.23% and 23.15%', while the conclusion states '11.23% and 21.15%' for density and flow. Please align these numbers with the actual entries in Table V.
  2. [Section VI-C3, Fig. 9 caption] The caption says 'initial sloth probability of MA γ0', but the experiments vary the MP's initial sloth probability; this should be corrected to avoid confusion.
  3. [Section VI-C2] The phrase 'rendering the property of γ and ε consistent diminishing be invalid' is ungrammatical and should be rewritten for clarity.
  4. [Section II-B] The text refers to a 'gradient-based shapely value approach'; the correct spelling is 'Shapley value'.
  5. [Section IV-B] The text mentions 'holomorphic encryption'; the standard term in this context is 'homomorphic encryption'.

Circularity Check

2 steps flagged · score 6.0 of 10

Incentive-mechanism convergence guarantee assumes terminal condition (44); Fig. 9's 'theoretical bound' is computed by imposing the equality the proof is supposed to establish.

  1. self definitional [Section V-B, Eqs. (38)-(50), Proposition 1 and its proof]
    "With the assumption that at least one condition in (44) is reached in the i∗-th settlement cycle, we can derive Proposition 1... Subsequently, upon the definition of γi∗ and (44), we can deduce γi∗−1ϵi∗−1ρ(β−1) > W, γi∗−1 > 0, ϵi∗−1 > 0. (50)"

    Convergence to 0 is stated as 'keep diminishing ... unless' conditions (44) are reached; the proof then assumes a condition of (44) is reached at i∗ and uses it to 'deduce' the strict inequalities (50). From the recurrence (38), 1/γi+1 = 1/γi + [γiϵiρ(β−1)−W]/S, so γ strictly decreases only while γiϵiρ(β−1)>W. The paper provides no argument that this bracket is positive before i∗ or that any branch of (44) is attained. The upper bound (45) is thus a restatement of the assumed terminal condition, not a derivation of it; the claimed convergence guarantee imports its own conclusion.

  2. fitted input called prediction [Section VI-C3, Fig. 9 and surrounding text]
    "Here, we consider that the theoretical upper bound γi∗ϵi∗ is calculated as per the first inequality of (50), i.e., γi∗ϵi∗ρ(β−1)=W."

    The 'theoretical upper bound' plotted as the validation of Propositions 1-2 is not obtained from the proof; it is computed by imposing the equality γi∗ϵi∗ρ(β−1)=W, which is the first terminal branch of (44). This is exactly the equality the proof was supposed to establish. Consequently, the probability gap γi∗ϵi∗−γϵ in Fig. 9 contains a term fixed by assumption, so the comparison to simulation only shows that the simulation is consistent with the assumed boundary, not that the bound is derived. The proof step is circular.

full rationale

The data-provider-selection component is self-contained: the MI model is trained on real drone-imaged traffic data and evaluated against external benchmarks (Central, Random, Oracle, VFLFS), so its empirical comparison is not circular. The only self-citation ([6], Zhu Han et al.) is background and does not carry any load-bearing argument. The circularity is confined to the incentive-mechanism theory. The paper claims γ and ε 'keep diminishing' from recurrences (38) and (43), but from (38) decrease requires γiϵiρ(β−1)>W; the proof asserts this as (50) 'upon the definition of γi∗ and (44)' without proving that any terminal condition in (44) is ever reached. Propositions 1 and 2 then use the assumed terminal index i∗ to state upper bounds whose numerators contain the same bracket. In Fig. 9, the 'theoretical upper bound' is explicitly calculated by setting γi∗ϵi∗ρ(β−1)=W, i.e., by assuming the equality the proof was meant to establish. The convergence-to-zero guarantee and the plotted bound are therefore equivalent to an unproven premise rather than derived consequences. This warrants a partial-circularity score of 6. Separately, the claimed accuracy gains (11.23%/21.15%) do not match Table V (11.26%/17.07%); that inconsistency is a correctness issue, not a circularity issue.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The framework introduces no new physical entities. The free parameters are the economic and cost-model constants chosen from references, plus simulation initial conditions. The main axioms are the transferability of the MI model, the ground-truth quality of drone labels, and the Markov assumption in the game, the last of which is also a derivation gap.

free parameters (8)
  • Reward W = $150
    Chosen in Table III as the payment from MA to MP per settlement cycle; drives equilibrium γ* and ε*.
  • Supervision cost S = $300
    Chosen in Table III as MA cost for inspecting MP data; drives γ*.
  • Base penalty ρ = $250 (default)
    Chosen in Section VI-C2; affects equilibrium probabilities and utility gains.
  • Additional penalty coefficient β = 11 (default)
    Chosen in Section VI-C2; moderates the penalty escalation.
  • Initial sloth probability γ0 = 0.3, 0.5, 0.7
    Simulation initial conditions for MP behavior.
  • Initial inspection probability ε0 = 0.3, 0.5, 0.7
    Simulation initial conditions for MA behavior.
  • MI sampling count = 50
    Number of samples drawn to approximate MI in Eq. (20).
  • Cost model constants from Table III = B=10MHz, p=0.1W, d(t)=100-500m, N0=-174dBm/Hz, D=80B, Δt=0.4s, η=1e-26, c=1.5e4, f_c=1e9Hz, E_r=7.2e-4J, τ_e=2.44e-4…
    Chosen with reference to [20]-[24] and used to compute the utility models; not measured in this paper.
assumptions (7)
  • standard math Donsker-Varadhan representation of KL divergence
    Used in Eq. (16) to express mutual information as a supremum over functions F; this is a standard mathematical tool.
  • domain assumption Drone imaging provides accurate ground-truth labels for all road segments
    Section IV-B: the MA trains the MI model on drone-collected traffic data and uses those states as labels; noisy or biased drone states would bias provider ranking.
  • domain assumption MI model trained on drone data transfers to MP trajectory data
    Section IV-B and Eq. (20): the same MI model F* is applied to each MP's data X_{k,n}; no distribution-shift correction is provided.
  • domain assumption Utility cost models from [20]-[24] apply to this VFL setting
    Section III-B uses energy and processing cost formulas from cited references; the constants are not measured here.
  • ad hoc to paper π' = 0, i.e., lazy MPs yield zero MA profit
    Section V-A states "we assume π' = 0 with well-designed TSE performance threshold"; this is necessary for the equilibrium calculation.
  • domain assumption Each MP contributes equally (1/K) to MA utility
    Eq. (9) assumes missing any MP's segment degrades the TSE model equally; this is a modeling choice.
  • domain assumption Markov assumption for the repeated game
    Section V-B says current strategies depend only on the previous settlement cycle and treats each cycle as the first interaction when deriving γ_i and ε_i; this is inconsistent with iterating the recurrence in Proposition 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Reliable Vertical Federated Learning Framework for Traffic State Estimation with Data Selection and Incentive Mechanisms." pith.science (2026). https://pith.science/paper/624Y3KER

@misc{pith2026250601285,
  author       = {Pith},
  title        = {Pith review of: A Reliable Vertical Federated Learning Framework for Traffic State Estimation with Data Selection and Incentive Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/624Y3KER}},
  note         = {Machine review of arXiv:2506.01285}
}
abstract

Vertical Federated Learning (VFL)-based Traffic State Estimation (TSE) offers a promising approach for integrating vertically distributed traffic data from municipal authorities (MA) and mobility providers (MP) while safeguarding privacy. However, given the variations in MPs' data collection capabilities and the potential for MPs to underperform in data provision, we propose a reliable VFL-based TSE framework that ensures model reliability during training and operation. The proposed framework comprises two components: data provider selection and incentive mechanism design. Data provider selection is conducted in three stages to identify the most qualified MPs for VFL model training with the MA. First, the MA partitions the transportation network into road segments. Then, a mutual information (MI) model is trained for each segment to capture the relationship between data and labels. Finally, using a sampling strategy and the MI model, the MA assesses each MP's competence in data provision and selects the most qualified MP for each segment. For the incentive mechanism design, given the MA can leverage the MI mode to inspect the data quality of MP, we formulate the interaction between MA and MP as a supervision game model. Upon this, we devise a penalty-based incentive mechanism to inhibit the lazy probability of MP, thereby guaranteeing the utility of MA. Numerical simulation on real-world datasets showcased that our proposed framework augments the traffic flow and density prediction accuracy by 11.23\% and 23.15\% and elevates the utility of MA by 130$\sim$400\$ compared to the benchmark.

Figures

Figures reproduced from arXiv: 2506.01285 by the authors.

Figure 1
Figure 1. Framework overview of the proposed VFL-based TSE with data provider selection and the incentive mechanism design. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. MAE comparison results versus varying the data quality [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Training curve of VFL-based TSE model among our [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figures from the paper (5 more)
Figure 7
Figure 7. Figure 7: The inspection probability of MA ϵ versus the base penalty ρ, where the initial inspection probability of MA of (a), (b), and (c) are all set as 0.5 while the initial sloth probability of MP is set as 0.3, 0.5, and 0.7. (a) (b) (c) 0 500 1000 1500 2000 2500 3000 0.0 0.…
Figure 8
Figure 8. Figure 8: The sloth probability of MP γ versus the base penalty ρ, where the initial inspection probability of MA of (a), (b), and (c) are all set as 0.5 while the initial sloth probability of MP is set as 0.3, 0.5, and 0.7. 150 175 200 225 250 0.00 0.02 0.04 0.06 0.08 0.10 0.1 …
Figure 9
Figure 9. Figure 9: The probability gap between theoretical upper bound [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: The comparison results in terms of MA’s inspection [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 13
Figure 13. Figure 13: The comparison results with varied ϵ0 and fixed γ0 = 0.5, where (a) is the averaged utility of MA, (b) is the averaged utility of MP, and (c) is the averaged social welfare. initial stage of the game spurs the MA to conduct supervision even when the MP’s probability o…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    Ising-traffic: Using ising machine learning to predict traffic congestion under uncertainty,

    Z. Pan, A. Sharma, J. Y .-C. Hu, Z. Liu, A. Li, H. Liu, M. Huang, and T. Geng, “Ising-traffic: Using ising machine learning to predict traffic congestion under uncertainty,” inProceedings of the AAAI Conference on Artificial Intelligence, Washington DC, Feb 2023

    work page 2023

  2. [2]

    A traffic flow dependency and dynamics based deep learning aided approach for network-wide traffic speed propagation prediction,

    H. Yang, L. Du, G. Zhang, and T. Ma, “A traffic flow dependency and dynamics based deep learning aided approach for network-wide traffic speed propagation prediction,”Transportation Research Part B: Methodological, vol. 167, pp. 99–117, Jan. 2023

    work page 2023

  3. [3]

    Pigat: Physics-informed graph attention transformer for air traffic state prediction,

    Q. Xu, Y . Pang, X. Zhou, and Y . Liu, “Pigat: Physics-informed graph attention transformer for air traffic state prediction,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 7, pp. 12 561–12 577, Sep. 2024

    work page 2024

  4. [4]

    Privacy-preserving data fusion for traffic state estimation: A vertical federated learning approach,

    Q. Wang and K. Yang, “Privacy-preserving data fusion for traffic state estimation: A vertical federated learning approach,”Transportation Research Part C: Emerging Technologies, vol. 168, p. 104743, Nov. 2024

    work page 2024

  5. [5]

    Optimal privacy control for transport network data sharing,

    B. Y . He and J. Y . Chow, “Optimal privacy control for transport network data sharing,”Transportation Research Part C: Emerging Technologies, vol. 113, no. 1, pp. 792–811, Apr. 2020

    work page 2020

  6. [6]

    Federated learning in vehicular edge computing: A selective model aggregation approach,

    D. Ye, R. Yu, M. Pan, and Z. Han, “Federated learning in vehicular edge computing: A selective model aggregation approach,”IEEE Access, vol. 8, pp. 23 920–23 935, Jan. 2020

    work page 2020

  7. [7]

    Vertical federated learning: Concepts, advances, and challenges,

    Y . Liu, Y . Kang, T. Zou, Y . Pu, Y . He, X. Ye, Y . Ouyang, Y .-Q. Zhang, and Q. Yang, “Vertical federated learning: Concepts, advances, and challenges,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 7, pp. 3615–3634, Jul. 2024

    work page 2024

  8. [8]

    A novel framework for traffic congestion management at intersections using federated learning and vertical partitioning,

    A. Chougule, V . Chamola, V . Hassija, P. Gupta, and F. R. Yu, “A novel framework for traffic congestion management at intersections using federated learning and vertical partitioning,”IEEE Transactions on Consumer Electronics, vol. 70, no. 1, pp. 1725–1735, Feb. 2024

    work page 2024

Show all 31 references
  1. [9]

    A mobility forecasting framework with vertical federated learning,

    F. Z. Errounda and Y . Liu, “A mobility forecasting framework with vertical federated learning,” inIEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC), Torino, Italy, Jun 2022

    2022

  2. [10]

    Fedsdg- fs: Efficient and secure feature selection for vertical federated learning,

    A. Li, H. Peng, L. Zhang, J. Huang, Q. Guo, H. Yu, and Y . Liu, “Fedsdg- fs: Efficient and secure feature selection for vertical federated learning,” inIEEE Conference on Computer Communications, New York, NY , May 2023

    2023

  3. [11]

    Vf-ps: How to select important participants in vertical federated learning, efficiently and securely?

    J. Jiang, L. Burkhalter, F. Fu, B. Ding, B. Du, A. Hithnawi, B. Li, and C. Zhang, “Vf-ps: How to select important participants in vertical federated learning, efficiently and securely?” inAdvances in Neural Information Processing Systems, New Orleans, LA, Nov 2022

    2022

  4. [12]

    Secure feature selection for vertical federated learning in ehealth systems,

    R. Zhang, H. Li, M. Hao, H. Chen, and Y . Zhang, “Secure feature selection for vertical federated learning in ehealth systems,” inIEEE International Conference on Communications, Seoul, South Korea, May 2022

    2022

  5. [13]

    Vertical federated learning-based feature selection with non- overlapping sample utilization,

    S. Feng, “Vertical federated learning-based feature selection with non- overlapping sample utilization,”Expert Systems with Applications, vol. 208, p. 118097, Dec. 2022

    2022

  6. [14]

    Less-vfl: Communication-efficient feature selection for vertical federated learning,

    T. Castiglia, Y . Zhou, S. Wang, S. Kadhe, N. Baracaldo, and S. Pat- terson, “Less-vfl: Communication-efficient feature selection for vertical federated learning,” inInternational Conference on Machine Learning, Hawaii, HI, Jul 2023

    2023

  7. [15]

    Coalitional federated learning: Improving communication and training on non-iid data with selfish clients,

    S. Arisdakessian, O. A. Wahab, A. Mourad, and H. Otrok, “Coalitional federated learning: Improving communication and training on non-iid data with selfish clients,”IEEE Transactions on Services Computing, vol. 16, no. 4, pp. 2462–2476, Jul-Aug. 2023

    2023

  8. [16]

    Chiron: A robustness-aware incentive scheme for edge learning via hierarchical reinforcement learning,

    Y . Liu, S. Guo, Y . Zhan, L. Wu, Z. Hong, and Q. Zhou, “Chiron: A robustness-aware incentive scheme for edge learning via hierarchical reinforcement learning,”IEEE Transactions on Mobile Computing, vol. 23, no. 8, pp. 8508–8524, Aug. 2024

    2024

  9. [17]

    Rate: Game-theoretic design of sustainable incentive mechanism for federated learning,

    B. Li, J. Lu, S. Cao, L. Hu, Q. Dai, S. Yang, and Z. Ye, “Rate: Game-theoretic design of sustainable incentive mechanism for federated learning,”IEEE Internet of Things Journal, Oct. 2024, early Access

    2024

  10. [18]

    Towards fair graph federated learning via incentive mechanisms,

    C. Pan, J. Xu, Y . Yu, Z. Yang, Q. Wu, C. Wang, L. Chen, and Y . Yang, “Towards fair graph federated learning via incentive mechanisms,” in Proceedings of the AAAI Conference on Artificial Intelligence, Vancou- ver, Canada, Feb 2024

    2024

  11. [19]

    Context-aware con- sensus algorithm for blockchain-empowered federated learning,

    Y . Zhao, Y . Qu, Y . Xiang, F. Chen, and L. Gao, “Context-aware con- sensus algorithm for blockchain-empowered federated learning,”IEEE Transactions on Cloud Computing, vol. 12, no. 2, pp. 491–503, Apr-Jun. 2024

    2024

  12. [20]

    An incentive mechanism of incorporating supervision game for federated learning in autonomous driving,

    Y . Fu, C. Li, F. R. Yu, T. H. Luan, and P. Zhao, “An incentive mechanism of incorporating supervision game for federated learning in autonomous driving,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 12, pp. 14 800–14 812, Dec. 2023

    2023

  13. [21]

    A novel contract theory-based incen- tive mechanism for cooperative task-offloading in electrical vehicular networks,

    S. A. Kazmi, T. N. Dang, I. Yaqoob, A. Manzoor, R. Hussain, A. Khan, C. S. Hong, and K. Salah, “A novel contract theory-based incen- tive mechanism for cooperative task-offloading in electrical vehicular networks,”IEEE Transactions on Intelligent Transportation Systems, vol. 2...

    2022

  14. [22]

    Intelligent edge computing in internet of vehicles: A joint computation offloading and caching solution,

    Z. Ning, K. Zhang, X. Wang, L. Guo, X. Hu, J. Huang, B. Hu, and R. Y . Kwok, “Intelligent edge computing in internet of vehicles: A joint computation offloading and caching solution,”IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 4, pp. 2212–2225, Apr. 2021

    2021

  15. [23]

    Dual-side optimization for cost-delay tradeoff in mobile edge computing,

    Y . Kim, J. Kwak, and S. Chong, “Dual-side optimization for cost-delay tradeoff in mobile edge computing,”IEEE Transactions on Vehicular Technology, vol. 67, no. 2, pp. 1765–1781, Feb. 2018

    2018

  16. [24]

    Hierarchical incentive mechanism design for federated machine learning in mobile networks,

    W. Y . B. Lim, Z. Xiong, C. Miao, D. Niyato, Q. Yang, C. Leung, and H. V . Poor, “Hierarchical incentive mechanism design for federated machine learning in mobile networks,”IEEE Internet of Things Journal, vol. 7, no. 10, pp. 9575–9588, Oct. 2020

    2020

  17. [25]

    Mutual information neural estimation,

    M. I. Belghazi, A. Baratin, S. Rajeshwar, S. Ozair, Y . Bengio, A. Courville, and D. Hjelm, “Mutual information neural estimation,” inInternational conference on machine learning, Stockholm, Sweden, Jul 2018

    2018

  18. [26]

    Vertimrf: Differentially private vertical federated data synthesis,

    F. Zhao, Z. Li, X. Ren, B. Ding, S. Yang, and Y . Li, “Vertimrf: Differentially private vertical federated data synthesis,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Barcelona, Spain, Aug 2024

    2024

  19. [27]

    Spatio-temporal graph convolutional net- works: A deep learning framework for traffic forecasting,

    B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional net- works: A deep learning framework for traffic forecasting,” inProceed- ings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, Stockholm, Sweden, Jul 2018

    2018

  20. [28]

    On the new era of urban traffic monitoring with massive drone data: The pneuma large-scale field experiment,

    E. Barmpounakis and N. Geroliminis, “On the new era of urban traffic monitoring with massive drone data: The pneuma large-scale field experiment,”Transportation Research Part C: Emerging Technologies, vol. 111, pp. 50–71, Feb. 2020

    2020

  21. [29]

    Hmm with non-emitting states for map matching,

    W. Meert and M. Verbeke, “Hmm with non-emitting states for map matching,” inEuropean Conference on Data Analysis (ECDA), Pader- born, Germany, Jul 2018

    2018

  22. [30]

    Vehicular blockchain- based collective learning for connected and autonomous vehicles,

    Y . Fu, F. R. Yu, C. Li, T. H. Luan, and Y . Zhang, “Vehicular blockchain- based collective learning for connected and autonomous vehicles,”IEEE wireless communications, vol. 27, no. 2, pp. 197–203, Apr. 2020. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 15 Zijun ...

    2020

  23. [2022]

    for contributions to game theory and distributed management of autonomous communication networks

    She is currently pursuing her Ph.D. degree in the department of architecture engineering at the Pennsylvania State University, University Park, PA, USA. Her current research interests include Build- ing Information Modeling, Blockchain, and Agentic Workflow. Daniel Mawunyo Doe...

    2018

Pith tools

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