Algorithmic fairness
by Oliver Thomas and Thomas Kehrenberg
Machine Learning
...using statistical techniques to give computer systems the ability to "learn" (e.g., progressively improve performance on a specific task) from data, without being explicitly programmed.
But there are problems:
Biased training data
Bias introduced by the ML algorithm
Let's have all the fairness!
If we are fair with regards to all notions of fair, then we're fair... right?
Independence based fairness (i.e. Statistical Parity)
$$ \hat{Y} \perp S $$Separation based fairness (i.e. Equalised Odds/Opportunity)
$$ \hat{Y} \perp S | Y $$For both to hold, then either $S \perp Y$, our data is fair, or $\hat{Y} \perp Y$, we have a random predictor.
Similarly, Sufficiency cannot hold with either notion of fairness.
So which fairness criteria should we use?
Consider a university, and we are in charge of administration!
We can only accept 50% of all applicants.
10,000 applicants are female and 10,000 of applicants are male.
We have been tasked with being fair with regard to gender.
We have an acceptance criteria that is highly predictive of success.
80% of those who meet the acceptance criteria will successfully graduate.
Only 10% of those who don't meet the acceptance criteria will successfully graduate.
As we're a good university we have a lot of applications from people who don't meet the acceptance criteria.
60% of female applicants meet the acceptance criteria.
40% of male applicants meet the acceptance criteria.
Remember, we can only accept 50% of all applicants
What should we do?
Truth Tables
Female Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | ||
| Don't Graduate |
Male Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | ||
| Don't Graduate |
How would we solve this problem being fair using Statistical Parity as our measure?
Female Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | 4000 | 1200 |
| Don't Graduate | 1000 | 3800 |
Male Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | 3300 | 500 |
| Don't Graduate | 1700 | 4500 |
How would we solve this problem being fair using Equal Opportunity as our measure?
Female Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | 4440 | 760 |
| Don't Graduate | 1110 | 3690 |
Male Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | 3245 | 555 |
| Don't Graduate | 1205 | 4995 |
How would we solve this problem being fair using Calibration by Group as our measure?
Female Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | 4800 | 400 |
| Don't Graduate | 1200 | 3600 |
Male Applicants
| Accepted | Not | |
|---|---|---|
| Actually Graduate | 3200 | 600 |
| Don't Graduate | 800 | 5400 |
Which fairness criteria should we use?
There's no right answer, all the above are "fair". It's important to consult domain experts to find which is the best fit for each problem. There is no one-size fits all.
Problems with doing this?
Any Ideas?
Problems with doing this?
What does this representation mean?
The learned representation is uninterpretable by default. Recently Quadrianto et al constrained the representation to be in the same same as the input so that we could look at what changed
Problems with doing this?
What if the vendor data user decides to be fair as well?
Referred to as "fair pipelines". Work has only just begun exploring these. Current research shows that these don't work (at the moment!)