Timezone: »
Data that is gathered adaptively --- via bandit algorithms, for example --- exhibits bias. This is true both when gathering simple numeric valued data --- the empirical means kept track of by stochastic bandit algorithms are biased downwards --- and when gathering more complicated data --- running hypothesis tests on complex data gathered via contextual bandit algorithms leads to false discovery. In this paper, we show that this problem is mitigated if the data collection procedure is differentially private. This lets us both bound the bias of simple numeric valued quantities (like the empirical means of stochastic bandit algorithms), and correct the p-values of hypothesis tests run on the adaptively gathered data. Moreover, there exist differentially private bandit algorithms with near optimal regret bounds: we apply existing theorems in the simple stochastic case, and give a new analysis for linear contextual bandits. We complement our theoretical results with experiments validating our theory.
Author Information
Seth Neel (University of Pennsylvania)
PhD student at University of Pennsylvania
Aaron Roth (University of Pennsylvania)
Related Events (a corresponding poster, oral, or spotlight)
-
2018 Oral: Mitigating Bias in Adaptive Data Gathering via Differential Privacy »
Wed. Jul 11th 02:20 -- 02:30 PM Room A6
More from the Same Authors
-
2021 : Adaptive Machine Unlearning »
Varun Gupta · Christopher Jung · Seth Neel · Aaron Roth · Saeed Sharifi-Malvajerdi · Chris Waites -
2022 : Individually Fair Learning with One-Sided Feedback »
Yahav Bechavod · Aaron Roth -
2022 : Individually Fair Learning with One-Sided Feedback »
Yahav Bechavod · Aaron Roth -
2023 Poster: Characterizing Multicalibration via Property Elicitation »
Georgy Noarov · Aaron Roth -
2023 Poster: Individually Fair Learning with One-Sided Feedback »
Yahav Bechavod · Aaron Roth -
2023 Poster: Multicalibration as Boosting for Regression »
Ira Globus-Harris · Declan Harrison · Michael Kearns · Aaron Roth · Jessica Sorrell -
2023 Oral: Multicalibration as Boosting for Regression »
Ira Globus-Harris · Declan Harrison · Michael Kearns · Aaron Roth · Jessica Sorrell -
2021 Poster: Differentially Private Query Release Through Adaptive Projection »
Sergul Aydore · William Brown · Michael Kearns · Krishnaram Kenthapadi · Luca Melis · Aaron Roth · Ankit Siva -
2021 Oral: Differentially Private Query Release Through Adaptive Projection »
Sergul Aydore · William Brown · Michael Kearns · Krishnaram Kenthapadi · Luca Melis · Aaron Roth · Ankit Siva -
2019 Poster: Differentially Private Fair Learning »
Matthew Jagielski · Michael Kearns · Jieming Mao · Alina Oprea · Aaron Roth · Saeed Sharifi-Malvajerdi · Jonathan Ullman -
2019 Oral: Differentially Private Fair Learning »
Matthew Jagielski · Michael Kearns · Jieming Mao · Alina Oprea · Aaron Roth · Saeed Sharifi-Malvajerdi · Jonathan Ullman -
2018 Poster: Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness »
Michael Kearns · Seth Neel · Aaron Roth · Steven Wu -
2018 Oral: Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness »
Michael Kearns · Seth Neel · Aaron Roth · Steven Wu -
2017 Poster: Meritocratic Fairness for Cross-Population Selection »
Michael Kearns · Aaron Roth · Steven Wu -
2017 Talk: Meritocratic Fairness for Cross-Population Selection »
Michael Kearns · Aaron Roth · Steven Wu -
2017 Poster: Fairness in Reinforcement Learning »
Shahin Jabbari · Matthew Joseph · Michael Kearns · Jamie Morgenstern · Aaron Roth -
2017 Talk: Fairness in Reinforcement Learning »
Shahin Jabbari · Matthew Joseph · Michael Kearns · Jamie Morgenstern · Aaron Roth