In my third post about the ethics of data science I’m heading into more challenging waters: fairness.
I was pointed to the work of Sorelle Friedler (thank you @natematias, @otfrom, and @mia_out) on trying to remove unfairness in algorithms by addressing the data that goes into them rather than trying to understand the algorithm itself.
I think this approach has some really positive attributes:
- It understands the importance of the data
- It recognises that the real world is biased
- It deals with the challenges of complex algorithms that may not be as amenable to interpretation as the linear model in my previous example.
Broadly speaking – if I’ve understood the approach correctly – the idea is this…
Rather that trying to interpret an algorithm, let’s see if the input data could be encoding for bias. In an ideal world I would remove the variables gender and ethnicity (for example) and build my model without them.
However, as we know, in the real world there are lots of variables that are very effective proxies for these variables. For example, height would be a pretty good start if I was trying to predict gender…
And so that is exactly what they do! They use the independent variables in the model to see if you can classify the gender (or sexuality, or whatever) of the record.
If you can classify the gender then the more challenging part of the work begins: correcting the data set.
This involves ‘repairing’ the data, but in a way that preserves the ranking of the variable… and the challenge is to do this in a way that minimises the loss of information.
It’s really interesting, and well worth a read.
Some potential difficulties
Whilst I think it’s a good approach, and may be useful in some cases, I think that there are some challenges that this approach needs to address, both at a technical, and at a broader level. Firstly though let’s deal with a couple of obvious ones:
- This approach is focused around a US legal definition of disparate impact. That has implications on the approach taken
- The concept of disparate impact is itself a contentious ethical position, with arguments for and against it
- Because the approach is based on a legal situation, it doesn’t necessarily deal with wider ethical issues.
As always, the joy of technical challenges are that you can find technical solutions. So here we go:
- The focus of the work has been on classifiers – where there is a binary outcome. But in reality we’re entering the world of probability, where decisions aren’t clear cut. This is particularly important when considering how to measure the bias. Where do you put the cutoff?
- Non-linear and other complex models also tend to work differentially well in different parts of the problem space. If you’re using non-linear models to determine if data is biased then you may have a model that passes because the average discrimination is fair (i.e below your threshold) but where there are still pockets of discrimination.
- The effect of sampling is important, not least because some discrimination happens to groups who are very much in the minority. We need to think carefully about how to cope with groups that are (in data terms) significantly under-represented.
- What happens if you haven’t recorded the protected characteristic in the first place? Maybe because you can’t (iPhone data generally won’t have this, for example), or maybe because you didn’t want to be accused of the bias that you’re now trying to remove. There is also the need to be aware of the biases with which this data itself is recorded…
The real difficulties
But smart people can think through approaches to those. What about the bigger challenges?
Worse outputs have an ethical dimension too:
If you use this approach you get worse outputs. Your model will be less accurate. I would argue that when considering this approach you also need to consider the ethical impact of a less predictive model. For example, if you were assessing credit worthiness then you may end up offering loans to people who are not going to be able to repay them (which adversely effects them as well as the bank!), and not offering loans to people who need them (because your pool of money to lend is limited). This is partially covered in the idea of the ‘business necessary defence’ in US law, but when you start dealing with probabilities it becomes much more challenging. The authors do have the idea of partially adjusting the input data, so that you limit the impact of correcting the data, but I’m not sure I’m happy with this – it smacks a bit of being a little bit pregnant.
Multiple protected categories create greater problems:
Who decides what protected categories are relevant? And how do you deal with all of them?
The wrong algorithm?
Just because one algorithm can classify gender from the data doesn’t mean that a different one will predict using gender. We could be discarding excellent and discrimination free models because we fear it might discriminate, rather than because it does. This is particularly important as often the model will be used to support current decision making, which may be more biased than the model that we want to use… We run the risk of entrenching existing discrimination because we’re worried about something that may not be discriminatory at all (or at least less discriminatory).
If it sounds like I think this approach is a bad one, let’s be clear, I don’t. I think it’s an imaginative and exciting addition to the discussion.
I like its focus on the data, rather than the algorithm.
But, I think that it shouldn’t be taken in isolation – which goes back to my main thesis (oh, that sounds grand) that ethical decisions need to be taken at all points in the analysis process, not just one.