Suggest edit — Algorithmic Fairness

Title

Name

Note

---
title: "Algorithmic Fairness"
type: concept
related: [Algorithmic Fairness]
source: https://www.jemoka.com/posts/kbhalgorithmic_fairness/
confidence: high
status: active
---

Could algorithms make more equitable decisions? No: algorithms may propagate and amplify biases&mdash;its not enough just to learn/optimize.
Collaboration with other fields has both a language gap and a value gap.
Technical Algorithmic Fairness fair scheduling, distributed computing, envy-freeness, cake cutting, stable matching, etc.
Interpreting Individual Probabilities &ldquo;I have n% probability of developing this thing&rdquo; &mdash; what does this mean? what does it capture in the environment?
Applications
insurance: how does actuarial scenarios work for this case&mdash;it is, in those cases, context dependent satisfying fairness maybe mis-generalized: discrimination can be subtle for a given measure (do you know it when you see it?) Definitions of fairness, however, is important to characterize a system.
Group Fairness Intuition: for a few protected groups \(S\), make sure that your predictor &ldquo;behaves similarly&rdquo; on \(S\) as on a general population \(U\)
&ldquo;similarly&rdquo; statistical parity: every prediction outcome i equally is as likely on \(S\) and \(U\) balance: similar false positive and false negative on both \(S\) and \(U\) calibration: prediction values are accurate on average on \(S\) and \(U\) these systems are all at odds with each other, and also are often at odds with the overall utility.
subgroups trying to advertise a burger store to vegetarians isn&rsquo;t going to work; so, fairness requires identifying subgroups \(s \subseteq S\) which are relevant to the task
multi-group fairness: offering &ldquo;fairness protection&rdquo; to every large subset of the population that can be identified given the data and computation limitations&mdash;fair by exhaustively protecting every group