Strategyproof Classification Under Constant Hypotheses: Mechanisms for Incentivizing Truthful Reporting
1. Reshef Meir, Ariel D. Procaccia, and Jeffrey S. Rosenschein Strategyproof Classification Under Constant Hypotheses: A Tale of Two Functions
2. Outline A very simple example of mechanism design in a decision making setting 8 slides An investigation of incentives in a general machine learning setting 2 slides
3. Motivation ECB makes Yes/no decisions at European level Decisions based on reports from national banks National bankers gather positive/negative data from local institutions Bankers might misreport their data in order to sway the central decision
4. A simple setting Set of n agents Agent i controls points Xi = {xi1,xi2,...} X For each xikXi agent i has a label yik{,} Agent i reports labels y’i1,y’i2,... Mechanism receives reported labels and outputs c+(constant ) or c(constant ) Risk of i: Ri(c) = |{k: c(xik) yik}| Global risk: R(c) = |{i,k: c(xik) yik}| = iRi(c)
6. Risk Minimization If all agents report truthfully, choose concept that minimizes global risk Risk Minimization is not strategyproof: agents can benefit by lying
8. Strategyproof approximation mechanisms VCG works (but is not interesting). Mechanism gives -approximation if returns concept with risk at most times optimal Mechanism 1: Define i as positive if has majority of + labels, negative otherwise If at least half the points belong to positive agents return c+ , otherwise return c- Theorem: Mechanism 1 is a 3-approx group strategyproof mechanism Theorem: No (deterministic) SP mechanism achieves an approx ratio better than 3
10. Randomized SP mechanisms Theorem: There is a randomized group SP 2-approximation mechanism Theorem: No randomized SP mechanism achieves an approx ratio better than 2
11. Reminder A very simple example of mechanism design in a decision making setting 8 slides An investigation of incentives in a general machine learning setting 2 slides
12. A learning-theoretic setting Each agent assigns a label to every point of X. Each agent holds a distribution over X Ri(c) = prob. of point being mislabeled according to agent’s distribution R(c) = average individual risk Each agent’s distribution is sampled, sample labeled by the agent Theorem: Possible to achieve almost 2-approximation in expectation under rationality assumption
13. Towards a theory of incentives in machine learning Classification: Richer concept classes Currently have strong results for linear threshold functions over the real line Other machine learning models Regression learning [Dekel, Fischer, and Procaccia, in SODA 2008]