How Machines Learn ​to Discriminate:

Embedding Fairness ​into Machine Learning ​Algorithms


By: Milad Farazian

Inspired by CSE 146: Ethics and Algorithms with Lise Getoor

What is Fairness in Machine Learning?

Fairness can be defined in many ways. For our purposes, we’ll focus on two:

Individual Fairness

  • People are treated equally to one another despite varying protected traits


  • Individual fairness prioritizes equal treatment

Group Fairness

  • People are treated equitably to one another despite varying protected traits


  • Group fairness prioritizes equal outcome

Algorithmic Bias

Data collected from the ​real world has bias, ​since the real world is ​affected by aspects of ​systemic oppression.


Algorithms develop from ​training off of real world ​data.

When retrieved data is ​biased, algorithms inherit ​that bias.

NOTE:

Biased algorithms ​could then reinforce ​systems of oppression

The Lack of Transparency Problem

Input Data

Predicted Results

“Black Box”

The “Black Box” (our machine learning algorithm) lacks nuance in ​demonstrating WHY predicted results are what they are.

?

Transparency for WHO is ​designing these algorithms is ​important for accountability.

Data Bias in Action

Google Vision Object Classifier Holds Racial Bias [1]

Correcting Data Bias in the Process

Pre-Processing

  • Detect Bias
  • Remove Bias (where possible)
  • Add Additional Data Sampling
  • Embedding (strings → numbers)

Processing

  • Evaluate Rankings for Bias
  • Search for Inference


Post-Processing

  • Look at Causal Relations
  • Verify Results for Accuracy vs. Fairness

OTHER KEY POINTS:

Question Your Own Bias!


Add Fairness to ​Algorithms...

Adding Fairness to Algorithms:

Individual Fairness

Equal Treatment


All individuals are treated equally regardless ​of their group identity or other factors




  • c represents algorithmic decision-making
  • p represents those with protected ​attributes
  • r represents those without


[3]

An example of Equal ​Treatment is all workers ​being paid equally for the ​same part-time job, not ​taking account of race, ​sex, etc.


[2]

Adding Fairness to Algorithms:

Group Fairness (1/2)

Equal Opportunity


Given two people from different groups are both good candidates, ​the probability of being selected is the same.



  • L = 1 represents whether the candidate is accepted
  • D = 0 represents those that are qualified enough to receive a ​positive outcome
  • p represents those with protected attributes
  • r represents those without.


[3]

[2]

Adding Fairness to Algorithms:

Group Fairness (2/2)

Demographic Parity


The probability of getting accepted should be ​the same for protected and regular candidates.


Prob_p [L = 1] = Prob_r [L = 1]. [3]


Similar to Equal Opportunity, only now we are ​considering that two demographic have equal ​likelihoods.



KEY POINT:

There is no ONE correct ​way to define fairness

Thank You!

[1] Kaysar-Bril, N. (2020, April). Google apologizes after its vision AI produced racist results. AlgorithmWatch. Retrieved November 16, 2021, from https://algorithmwatch.org/en/google-vision-racism/.


[2] MobilizeGreen. (2021, February 10). Environmental equity vs. environmental justice: What's the difference? MobilizeGreen. Retrieved November 18, 2021, from https://www.mobilizegreen.org/blog/2018/9/30/environmental-equity-vs-environmental-justice-whats-the-difference.


[3] Fu, Aseri, M., Singh, P., & Srinivasan, K. (2021). “Un”Fair Machine Learning Algorithms. Management Science. https://doi.org/10.1287/mnsc.2021.4065