writing

Fresh Data is Fairer Data

September 2022

Establishing what is fair is hard, and machine learning models need fair exemplars to correctly learn to make decisions that we would agree with. The fresher the data the more likely it is to be fair (although it is likely still far from our desires).

Approaches to Fairness

First, let us establish a baseline level of fairness that we can formalise. Taking a leaf from the justice system, let us call a decision fair if the decision does not rely on certain factors such as “race” or “gender” 1.

Imagine we want to make a predictor Y^\hat{Y} that attempts to predict outcome YY, we can imagine it to be fair if it predicts YY without reliance on any protected characteristic 2 PP.

Now there are a few simple, but flawed, approaches we might consider to construct fair data from our unfair reality:

  1. Remove any sensitive attributes: If the model doesn’t have the data, then it can’t use it to make predictions, right? Well, no. There is “proxy discrimination” where the predictive power of a seemingly neutral characteristic is at least partially attributable to its correlation without our characteristic PP. As such we cannot rely on simply omitting these characteristics, and must consider fairness from a different perspective.
  2. Remove correlated attributes: “Aha! Well, you said it was correlated with PP, why don’t we just calculate the correlations and then remove any characteristic over a certain threshold.” Statistical measures of correlation can often result in two or more uncorrelated features being able to be combined to form a highly correlated one. Our filtering process is incomplete. Our morally objectionable prior 3 is likely redundantly encoded in the feature space.

So what is the approach?

Well, imagine a causal chain i.e. say GG is a potential cause of LL because there exists a causal chain GILG \rightarrow I \rightarrow L. These indirect links are tricky!

Using II to predict LL is fair only as long as its predictive power comes from II, and not its underlying relationship to GG. This seems completely intractable; say for example that GG is gender, II is excess income and LL is loan repayment. Income feels a “fair” feature to judge upon because the majority of the predictive power is from having a high income, not the applicant’s gender.

A way out?

Ultimately we’re trying to encode our societies’ moral compass into the model. It’s not an easy task.

This seems true anecdotally, with instances such as language models showing a decrease in gender bias over time

Re-training models with fresh data will neatly sidestep the whole problem of fairness, older data is more likely to be steeped in priors that we now disagree with. Much like the generations to come will look back on us and wonder how we could have been so unfair.

While I doubt fine-tuning would be sufficient, we may be able to integrate some historical down weight during training that allows the usage without all the negative consequences.

  1. This is not a foolproof approach - we could be biased towards an unprotected characteristic (e.g. blue eyes) that would certainly meet the human criteria for unfair decision-making. However, we must restrict our definition of fair to a minimum baseline, otherwise, we’ll be specifying the problem to such a degree that we do not need to learn a decision.

  2. There are several protected characteristics/examples of discrimination, for example, the UK Government lists: age, gender reassignment, being married or in a civil partnership, being pregnant or on maternity leave, disability, race including colour, nationality, ethnic or national origin, religion or belief, sex, and sexual orientation.

  3. I have a thing about the word bias, you can read about it if you’re interested.