Individual Fairness in Hindsight

on YouTube ↗31m

" summary

Defines individual fairness in hindsight as a Lipschitz-style rule that only imposes past-based lower bounds on current decisions so decisions can monotonically become more conducive but not worse, illustrated with a loan model (blue 1/4, green 3/4, allowed difference 0.5) where optimal steady policy is (0,0.5). Proposes CAFÉ (cautious fair exploration): start with epsilon loans, learn which type defaults, then increase loans monotonically. Gives regret rates (order log T when KL rates are favorable) and discusses legal and pricing analogies.

AI-written summary · about

" chapters (11)

1. 0:00Intro
2. 0:42Definitions of Fairness
3. 1:11Similarity = Lipschitzness
4. 3:30Fairness-across-time
5. 5:12Individual Fairness and Learning
6. 6:31Individual Fairness in Hindsight
7. 13:12Outline of the talk
8. 14:50Utility Maximization: Dynamic
9. 16:32Learnability of Utility
10. 17:10Cautious Fair Exploration (CaFE)
11. 25:13A fundamental tradeoff

" related talks

• 34mInterpretability and Algorithmic FairnessAlice Xiang · Interpretable Machine Learning in Natural and Social Sciences (2022)
• 31mPreference Informed FairnessGal Yona · Beyond Differential Privacy 2019
• 53mGroups and Individuals in Fairness and ProbabilityCynthia Dwork · Privacy and the Science of Data Analysis 2019
• 37mAlgorithmic Fairness From The Lens Of Causality And Information TheorySanghamitra Dutta · Algorithmic Aspects of Causal Inference 2022
• 41mLow rate loans for ladies, stags pay extra: The Role of Ethics in AI/MLChris Stucchio · 50p 2018
• 30mIndividual Statistical FairnessAaron Roth · Beyond Differential Privacy 2019
• 42mFairness in Machine Learning: From Theory to PracticeAlex Karsten · Open AI + Data Forum @ Open Source Summit North America 2023
• 30mBuilding fairer models for financeAndrew Weeks · PyData Global 2020