I started thinking about the fairness of machine learning models recently. Wiki page for Fairness_(machine_learning) defines fairness as:

In machine learning, a given algorithm is said to be fair, or to have fairness if its results are independent of some variables we consider to be sensitive and not related to it (f.e.: gender, ethnicity, sexual orientation, etc.).

UC Berkley CS 294 in turn defines fairness as:

understanding and mitigating discrimination based on sensitive characteristics, such as, gender, race, religion, physical ability, and sexual orientation

Many other resources, like Google in the ML Fairness limit the fairness to these aforementioned categories and no other categories are considered.

But fairness is a lot broader context than simply these few categories mentioned here, you could easily add another few like IQ, height, beauty anything that could have a real impact on your credit score, school application or job application. Some of these categories may not be popular in existing datasets nowadays but given the exponential growth of data, they will be soon, to the extent, that we will have an abundance of data about every individual with all their physical and mental categories mapped into the datasets.

Then the question would be how to define fairness given all these categories presented in the datasets. And will it even be possible to define fairness if all physical and mental dimensions are considered as it seems that when we do so, all our weights in, say, the neural nets should be exactly the same, i.e., giving no discriminator in any way or form towards or against any physical or mental category of a human being? That means that a machine learning system that is fair across all possible dimensions will have no way of distinguishing one human being from another which would render these machine learning models useless.

To wrap it up, while it does make perfect sense to withdraw bias towards any individual given the categories like gender, ethnicity, sexual orientation, etc., the set is not closed and with the increasing number of categories being added to this set, we will inevitably arrive at a point where no discrimination (in a statistical sense) would be possible.

And that's why my question, are fair machine learning models possible? Or perhaps, the only possible fair machine learning models are those that arbitrarily include some categories but ignore other categories which, of course, if far from being fair.

  • $\begingroup$ You appear to have deliberately misconstrued the definition of fairness in this question to always include all factors, making it impossible to answer without a frame challenge to your definition. Most people would not extend "fair" to always include all categories for all uses, and it results in a reductio ad absurdum argument against automated fairness which is unlikley to be the goal of any AI decision-making system. Is that the question you want to ask? $\endgroup$ – Neil Slater May 26 at 11:07
  • $\begingroup$ I see that you are not quite agreeing with how I phrased fairness. Well, these definitions, as you see were not coined by myself. I merely refer to them. Are they clear and concise. Clearly not. That's why i even highlighted the word etc. such as should also be highlighted.There is a huge room for manoeuvre with regard to which categories should or should not be treated fairly. If I am reading you correctly fair in AI should be include just some categories and exclude others? Is that your point? $\endgroup$ – matcheek May 26 at 11:55
  • $\begingroup$ My point is this: We cannot do philosophical debates on Stack Exchange. If your question resolves to "Why is fairness not usually considered all-inclusive of all traits?" then it may be better for Philosophy Stack Exchange. If your question is about how fairness is handled by AI systems, you may want to step back from your reductio ad absurdum (to paraphrase "In an absolutely fair system, no measurable trait may be used to make a decision, so what's the point?") because no-one treats fairness like that in practice $\endgroup$ – Neil Slater May 26 at 12:23
  • $\begingroup$ Or another way to put it: I would not want to post an answer briefly dismissing the over-broad definition of "fairness" you have used and start talking about fairness in AI systems, only to enter into a debate with you about the definition you have used. However, if you are looking for guidance/correction on the definition itself, it could be answered (perhaps better on Philosophy SE though) $\endgroup$ – Neil Slater May 26 at 12:26
  • 2
    $\begingroup$ @matcheek I would reformulate your question "are fair machine learning models possible?" as "Is there a definition of fairness applicable to ML models that takes into account all possible factors that can affect 'fairness'?" (or something like that). In general, I suggest you formulate your question so that it can be answered more objectively and that it avoids debates. $\endgroup$ – nbro May 26 at 12:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.