ML fairness gym

Vit Listik

13.2.2020

Motivation

• ML is doing high-impact decisions
• https://github.com/google/ml-fairness-gym/blob/master/papers/acm_fat_2020_fairness_is_not_static.pdf
• Google
• 2019
• https://github.com/google/ml-fairness-gym
• Simulating potential long-run impacts of deploying machine learning-based decision systems
• Using OpenAI Gym
• Testing static fairness (test set)

2

Lending problem

• 2 groups (observable)
• Lending money
• Credit score
• https://1.bp.blogspot.com/-o8-7POXOmgA/XjsWStg54NI/AAAAAAAAFQI/QiDLNWNfdlUrn79sieIXXHm5fDWVfd9OQCLcBGAsYHQ/s1600/image5.gif

3

Process

• 2 banks (agents) with 2 strategies
• Max reward
• Equal OP

4

5

Allocation problem

• agent is tasked with allocating finite units of attention across different sites with the goal of discovering harmful incidents
• Insufficient resources to cover everything

6

7

College admissions

• applicants are able to manipulate their scores (at a cost) to obtain a desirable decision from the agent

8

9

Resources

• https://ai.googleblog.com/2020/02/ml-fairness-gym-tool-for-exploring-long.html
• https://bair.berkeley.edu/blog/2018/05/17/delayed-impact/

10