Making Decisions under Outcome Performativity
summary
Defines performative omnipredictor: a regression p(x,y') that, after a simple post‑processing, simultaneously approximates performative optima for any loss set L and comparator class H; proves existence of p of circuit size poly(s,1/ε) when L and H have circuit size s. Builds a boosting‑style auditing algorithm that enforces performative outcome indistinguishability by iteratively finding distinguishers and updating p, and shows the audit step reduces to cost‑sensitive classification, yielding statistically and computationally efficient construction using polynomial samples and iterations.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:27Social Predictions are Performative
- 3:12A Brief Intellectual History...
- 4:48The Outcome Performativity Thesis
- 6:50The Performative Prediction Framework
- 8:12Ambiguity of Optimality
- 9:45Omniprediction in Supervised Learning
- 11:27Performative Omniprediction
- 15:01Omniprediction is not
- 15:38The Postprocessing: An Example
- 17:21The Performative Postprocessing
- 20:19Defining Indistinguishability
- 22:14Indistinguishability Algorithm
- 25:35Summing Up
related talks
39mA PAC Approach to Application-Specific Algorithm Selection
37mOmnipredictors
56mOn the Curses of Future and History in Off-policy Evaluation in non-Markov Environments
20mStatic Posterior Inference of Bayesian Probabilistic Programming via Polynomial Solving
45mAGT and Machine Learning
48mFrom Predictions to Decisions: Limitations and Possibilities of Optimization from Samples
36mThe Overlap Gap Property For Inference: Power And Limitations
45mFollow the Leader with Dropout Perturbations