Research map / Learning theory and algorithms / Fairness
Incentivizing Time-Aware Fairness in Data Sharing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes time-aware incentives and two reward schemes (time-aware reward cumulation and time-aware data valuation based on Shapley values) so parties joining earlier in data sharing receive higher-value rewards while preserving fairness and individual rationality.
Problem
Existing collaborative machine learning incentive frameworks assume all parties join simultaneously; when parties join at different times, fairness and individual rationality conflict with rewarding early contribution, and the Shapley value ignores joining times.
Approach
Define time-aware incentives (equal-time symmetry/desirability, time-based monotonicity and strict monotonicity, necessity) and two reward schemes: a weighted sum of Shapley values over time intervals, and a time-aware valuation function that discounts superadditive coalition dividends by exp(-gamma * joining time) before applying the Shapley value; rewards are realized via likelihood tempering or
Key results
- On the Friedman dataset with three parties, rewards decrease as party 1 joins later, but party 1 still receives reward above its individual data value (individual rationality holds).
- On Friedman, party 1 gets lower reward than party 2 when it joins at t1 >= 3 (beta=1, gamma=0.5), showing time-based desirability emphasis can be tuned.
- On MNIST, rewards decrease with later joining and the highest-reward party achieves the lowest MNLP, though the trend is not strictly monotonic due to subset-selection approximation and NN training randomness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use time-aware reward cumulation instead of time-agnostic Shapley rewards | Collaborative data sharing · Friedman dataset | reward value | Helped |
| Use time-aware data valuation instead of time-agnostic Shapley rewards | Collaborative data sharing · Friedman dataset | reward value | Helped |
| Delay party 1's joining time | Time-aware reward schemes · Friedman dataset | reward value | Hurt |
| Set β=0.7 and γ=1 instead of β=1 and γ=0.5 (ablation) | Time-aware reward schemes · Friedman dataset, party 1 versus party 2 | reward-value difference | Mixed |
| Increase β (ablation) | Time-aware reward cumulation · Friedman dataset | emphasis on early joining | Hurt |
| Increase γ (ablation) | Time-aware data valuation · Friedman dataset | emphasis on early joining | Helped |
| Use γ=0, removing joining-time influence from data valuation (ablation) | Time-aware data valuation · Friedman dataset | reward-value variation with joining time | Hurt |
| Delay party 1's joining time | Time-aware reward schemes · CaliH dataset | MNLP | Hurt |
| Delay party 1's joining time | Time-aware reward schemes with subset selection · MNIST | MNLP | Mixed |
| Use subset selection for reward realization (ablation) | Time-aware reward schemes · MNIST | MNLP | Mixed |
| Increase party 1's joining time while keeping its data value higher than party 2's | Time-aware reward schemes · Friedman dataset, party 1 versus party 2 | reward-value ranking: β=1 and γ=0.5: t1 ≥ 3; β=0.7 and γ=1: t1 ≥ 2 | Mixed |
| Delay party 1 while party 3 has substantially less valuable data | Time-aware reward schemes · Friedman dataset, party 1 versus party 3 | reward-value ranking | No clear effect |
Surprising findings
- Rewards of all parties drop when one party joins late, not only the late party, because others benefit from collaboration for a shorter duration.
- The trend of MNLP versus joining time on MNIST is not strictly monotonic.
Limitations
- Exact Shapley computation is exponential in the number of parties; scalability relies on approximation.
- One-time data sharing only; does not address online federated learning setting.
- Privacy risks of data sharing are not addressed.
Experimental setup
Datasets: Friedman synthetic dataset California housing (CaliH) MNIST
Benchmarks: Friedman California housing MNIST
Metrics: reward value (conditional information gain or validation acc mean negative log probability (MNLP)
Baselines: Shapley value time-weighted Shapley value
Builds on: Shapley value Harsanyi dividend conditional information gain likelihood tempering
Closest papers by meaning
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.86)
- Paid with Models: Optimal Contract Design for Collaborative Machine Learning (AAAI 2025 · similarity 0.85)
- A Cramér–von Mises Approach to Incentivizing Truthful Data Sharing (NeurIPS 2025 · similarity 0.85)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.85)
- Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors (NeurIPS 2024 · similarity 0.84)
- Temporal Fair Division (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.