핵심 개념
Individual fairness auditing schemes aim to balance predictive accuracy and fairness violations, enhancing online learning algorithms.
초록
The content delves into the concept of individual fairness in online learning, focusing on auditing schemes for detecting violations. It introduces a novel framework for auditing unfairness, presents oracle-efficient algorithms, and discusses the complexity of auditing. The study aims to achieve simultaneous no-regret guarantees for accuracy and fairness in online learning settings.
Introduction:
Increasing use of algorithms in critical decision-making domains.
Formalizing notions of fairness and proposing accurate algorithms.
Individual Fairness and Monotone Auditing Schemes:
Defining fairness violations and auditors' roles.
Introducing monotone aggregation functions for auditing schemes.
Online Learning with Individual Fairness:
Defining misclassification and unfairness losses.
Formulating learning objectives for simultaneous no-regret guarantees.
Partial Information:
Policies refraining from violations slightly below sensitivity level α.
Conclusion and Future Directions:
Challenges in achieving simultaneous accuracy-fairness objectives.
통계
Using our generalized framework, we present an oracle-efficient algorithm achieving an upper bound of O(√T) for regret...
Our construction will only require making ˜O(α−2) calls to an optimization oracle on every round...