Incentivizing Truthful Data Sharing in Collaborative Machine Learning and Data Marketplaces
summary
Mechanisms claim that truthful data submission is a (approximate) Nash equilibrium with truthful reporting and that, for normal mean estimation, the social penalty is at most twice the global optimum while each agent only needs ~1/√M of the solo sample count. A general, prior-agnostic reward uses a Cramér–von Mises-based scoring rule (feature-wise for images/text) that makes fabricating extra synthetic points increase loss, whereas submitting more real points (e.g. 500 → 2,500 sentences) lowers loss; applied to federated learning and data marketplaces with empirical success.
AI-written summary · about
chapters (11)
- 0:00Collaborative data sharing vision
- 1:05Platforms vs marketplaces
- 2:29Free-riding and fabrication risks
- 7:24Truthful data sharing approaches
- 10:55Mean estimation with agents
- 25:51General data spaces without post-processing
- 37:40Warm-up mechanism for 1D data
- 40:14Prior-agnostic loss function
- 42:27Experimental validation
- 45:01Applications in federated learning
- 46:50Impossibility results
related talks
01h 27mTutorial: Incentives for Collaborative Learning and Data Sharing, Part II
01h 00mTutorial: Incentives for Collaborative Learning and Data Sharing, Part Ib
45mTutorial: Incentives for Collaborative Learning and Data Sharing, Part I
47mNot all data are created Equal: Robust Mean Estimation from Heterogeneous and Unreliable Users
23mFalse Consensus, Information Theory, and Prediction Markets
46mEliciting and Enforcing Subjective Individual Fairness
14mFair And Reliable Machine Learning For High-Stakes Applications:approaches Using Information Theory
27mLearning and Strongly Truthful Multi-task Setting Peer Prediction: A Variational Approach