More Dominantly Truthful Multi-task Peer Prediction with a Finite Number of Tasks
summary
Introduces a new information measure, volume mutual information, that equals the determinant-based mutual information under uniform density and satisfies the data processing inequality so payments drop under any garbling. Uses polynomial densities so the measure has a polynomial form and admits unbiased estimation from a finite number D of tasks (D equals the polynomial degree). Shows the binary 2×2 case as a cube slice, constructs a family of dominant-truthful, prior-free multi-task peer-prediction mechanisms, and gives a polynomial approximation to a threshold payment scheme.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:15Multi-task peer prediction [Dasgupta and Ghosh 2013]
- 3:35Mutual Information Paradigm [Kong and Schoenebeck 2018]
- 5:44Data Processing Inequality = Strategy decreases payment
- 8:31Key challenge Infinite number of rating tasks
- 9:52Determinant based Mutual Information Kong 2020
- 15:19More...? Mechanism design = Information measure design
- 16:38Volume mutual information Counting the number of people that are less clever than you
- 23:32Visualization Distribution space (binary) Domain 2
- 26:52Volume mutual information Binary visualization
- 28:32Threshold payment scheme Approximating via polynomial VMI
- 31:01Conclusion And future work
related talks
50mStrong Truthfulness in Multi-signal Peer Prediction with Overlapping Tasks
50mDiversity in Matching Markets
32mDynamic Matching Models
29mClearing Matching Markets Efficiently: Informative Signals and Match Recommendations
44mClearing Matching Markets Efficiently: Informative signals and Match Recommendations
44mOnline Advertisements and Online Algorithms
25mBudget Pacing in Repeated Auctions: Regret and Efficiency without Convergence
42mPredictive Load-Balancing: Unfair but Faster & more Robust