Making machine learning algorithms fair is a tricky business.
What does “fair” even mean? You might think it’s easy to define but what’s fair to you might not be fair to someone else. Even defining the word is a slippery slope filled with pitfalls and traps.
Maybe you think it’s as simple as deleting age or race or gender from a data set but think again. In the real world making algorithms fair is a lot harder and it always leads to trade offs.
In this amazing post, called A Tutorial on Faireness in Machine Learning, by PhD student Ziyuan Zhong, we get a whirlwind tour through the different approaches to rooting out biases in algorithms.
Maybe you’re building a hiring system and you think that just deleting race or gender is enough, but it’s not because characteristics cascade into multiple other aspects of a candidates primary attributes, like their name, the groups they join, where they live and where they went to school.
Zhong breaks down six different approaches data scientists use to deal with these challenges in production models across the world:
- Demographic Parity
- Equalized Odds
- Predictive Rate Parity
- Individual Fairness
- Counterfactual fairness
While none of these methods are foolproof, knowing them inside and out can help you rip bias up by the roots.