Thread by @annagranta

A thread about the a levels fiasco and algorithmic bias. From an ex machine learning engineer with a passion for social justice. #alevels2020

An algorithm has been used to decide students a level results this year as a result of #COVID19

Is that algorithm fair or is it biased? The quality of an algorithm usually rests on the quality of the data used to train it. (If a totally inappropriate method is used there best data can't save you, but usually the data is the key factor).

Data scientists have a name for this 'garbage in, garbage out'

So what data could be used to determine an exam grade, and what data was used?

We could potentially use lots of attributes of the student: teacher predicated grades, students past results, student height, student gender....

Wait! Did I say gender?!! Surely that would be discriminatory?

Yes, using gender to determine someone's exam result would be discriminatory. We'd better take that off the list.

Hang on, if I tell you my height (4ft11) you could take a pretty good guess at my gender. In technical terms my height and my gender and not independent!

This means that to be totally fair we need to remove height too. Otherwise the computer can use my height too guess my gender and then base my grade on my likely gender. That would be biased!

Actually it turns out that all of the bits of data I mentioned are biased because none of them is independent of gender or of race. Since discrimination against someone based on a protected characteristic is not only wrong but also illegal what do we do?

Well, we could look for other data. The government used data about the pupils schools past performance. Is this unbiased?

No Gavin, it's biased. A schools performance might not tell you the gender of an individual student, but it is not independent of the racial makeup of the school.

Since schools racial makeup rarely changes much year to year, the schools performance is not independent of the pupils race.

So what can we do? Is it possible to come up with a non racist system for giving out #AlevelResults ?

Mathematically, the answer is yes! Randomly allocating grades is free of bias, as is giving all students the same grade. Mathematicians don't have a reputation for being practical.

So I see two options,
1) find all possible data on students, check how closely it corrallates with each protected characteristic and bin that data that corrallates more than some 'acceptable' amount.
2) give everyone an A.

Because I wasn't to propose a modification to 'garbage in, garbage out' (remember this is an important industry principal).

'biased garbage in, biased garbage out'. And I don't think biased garbage is good enough to determine a level results.

I used to work in a school with a very high proportion of disadvantaged students and the head said this:

These are the kids who have to get it right first time, because they don't get second chances.

The most disadvantaged children will not get a chance to hustle hard and fail upwards. They need us to fight for them to get this chance, because it could be the only one they have.

Please email your MP, tag a journalist or retweet. Biased Garbage grades, based on biased garbage data are not good enough.