Indirect Influence and Fairness in Machine Learning

Date
2020
Journal Title
Journal ISSN
Volume Title
Publisher
Producer
Director
Performer
Choreographer
Costume Designer
Music
Videographer
Lighting Designer
Set Designer
Crew Member
Funder
Rehearsal Director
Concert Coordinator
Moderator
Panelist
Alternative Title
Department
Haverford College. Department of Computer Science
Type
Thesis
Original Format
Running Time
File Format
Place of Publication
Date Span
Copyright Date
Award
Language
eng
Note
Table of Contents
Terms of Use
Rights Holder
Access Restrictions
Haverford users only until 2021-01-01, afterwards Tri-College users only
Tripod URL
Identifier
Abstract
Methods for measuring fairness in machine learning often operate by quantifying the relationship between some protected feature (e.g., race or gender) and the predictions of a model. When we are interested in understanding whether a model treats an individual fairly, it can be useful to understand whether the protected features associated with that individual affect the prediction they are given. However, a protected feature can influence the predictions of a model indirectly, through proxies: other features the model has access to which are related to the protected feature. Two main strategies have emerged for measuring indirect influence. One strategy generates a second copy of the data in which the protected information is removed from the rest of the data and observes how much worse a model performs on the altered data. Alternatively, other methods try to understand how features act as proxies for protected features and compute indirect influence without altering the input data. We discuss four recent methods for computing indirect influence and explore the strengths and limitations of each method. Lastly, we identify limitations of the current study of indirect influence and propose research directions to mitigate these limitations by precisely defining the objectives of indirect influence measurement in information theoretic terms.
Description
Citation
Collections