Multicalibration, Universal Adaptability and Causality
summary
Shows a concrete reduction from adapting a study to many unseen targets to enforcing multicalibration: under conditional independence (Y ⟂ Z | X) and mild coverage (propensity bounded away from zero), the construction produces a single-source estimator (a predictor released to targets, not just a number) that needs only unlabeled target samples and is comparable to propensity score reweighting for every target. Algorithmic tools: multicalibration/multi-accuracy, outcome-indistinguishability and omni-predictor ideas; experiments report similar or sometimes better accuracy and extensions to treatment-effect estimands.
AI-written summary · about
chapters (14)
- 0:00Intro
- 1:25LET'S TALK ABOUT IT
- 3:11GENERALIZATIONS FROM A SOURCE TO A TARGET
- 4:26STATISTICAL ESTIMATION SETUP
- 7:52KEY ASSUMPTIONS
- 10:57PROPENSITY SCORE = VALID INFERENCE
- 12:29FITTING THE PROPENSITY SCORE
- 13:15CHALLENGES WITH THIS PARADIGM
- 20:55ALGORITHMIC FAIRNESS
- 28:47MULTICALIBRATION UNIVERSAL ADAPTABILITY
- 33:07EXTENSIONS AND EXPERIMENTS
- 34:50PROPENSITY SCORING FOR TREATMENT EFFECT
- 37:32DISCRIMINATION BY NON-CAUSALITY
- 39:23MULTI-MODEL CAUSALITY!
related talks
01h 21mMulticalibration: Towards Fair Decision Making
12mA Complexity-Theoretic Perspective on Fairness
47mA Multi-Group Perspective on Fairness
37mOmnipredictors
19mHappyMap: A Generalized Multicalibration Method
48mOnline Adversarial Multicalibration And (Multi)Calibeating
24mLoss Minimization through the lens of Outcome Indistinguishability
44mCausal Reductions