[ewjordan]

If they were using any sort of neural networks approach with stochastic gradient descent, the network would have to spend some "gradient juice" to cut a divot that recognizes and penalizes women's colleges and the like. It wouldn't do this just because there were fewer women in the batches, rather it would just not assign any weight to those factors.

Unless they presented lots of unqualified resumes of people not in tech as part of the training, which seems like something someone might think reasonable. Then, the model would (correctly) determine that very few people coming from women's colleges are CS majors, and penalize them. However, I'd still expect a well built model to adjust so that if someone was a CS major, it would adjust accordingly and get rid of any default penalty for being at a particular college.

If the whole thing was hand-engineered, then of course all bets are off. It's hard to deal well with unbalanced classes, and as you mentioned, without knowing what their data looks like we can only speculate on what really happened.

But I will say this: this is not a general failure of ML, these sorts of problems can be avoided if you know what you're doing, unless your data is garbage.

[lalaland1125]

> It wouldn't do this just because there were fewer women in the batches, rather it would just not assign any weight to those factors.

That's exactly the issue we are talking about here. Woman's colleges would have less training data so they would get updated less. For many classes of models (such as neural networks with weight decay or common initialization schemes) this would encourage the model to be more "neutral" about women and assign predictions closer to 0.5 for them. This might not affect the overall accuracy for women (as it might not influence whether or not they go above or below 0.5), but it would cause the predictions for women to be less confident and thus have a lower ranking (closer to the middle of the pack as opposed to the top).

[ewjordan]

I don't think I'm with you. A neural net cannot do this - picking apart male and female tokens requires a signal in the gradients that force the two classes apart. If there's no gradient, then something like weight decay will just zero out the weights for the "gender" feature, even if it's there to begin with. Confidence wouldn't enter in, because the feature is irrelevant to the loss function.

A class imbalance doesn't change that: if there's no gradient to follow, then the class in question will be strictly ignored unless you've somehow forced the model to pay attention to it in the architecture (which is possible, but would take some specific effort).

What I'm suggesting is that it's likely that they did (perhaps accidentally?) let a loss gradient between the classes slip into their data, because they had a whole bunch of female resumes that were from people not in tech. That would explain the difference, whereas at least with NNs, simply having imbalanced classes would not.

[8note]

supposing waiter and waitress are both equally qualifying for a job, and most applicants are men, won't the ai score waiter as being more valuable than waitress?

[Reelin]

Not generally. The entire point being made is that whether one feature is deemed to be more valuable than another feature depends not just on the data fed into the system but also on the training method used.

Specifically, the gp is pointing out that typical approaches will not pay attention to a feature that doesn't have many data points associated with it. In other words, if it hasn't seen very much of something then it won't "form an opinion" about it and thus the other features will be the ones determining the output value.

Additionally, the gp also points out that if you were to accidentally do something (say, feed in non-tech resumes) that exposed your model to an otherwise missing feature (say, predominantly female hobbies or women's colleges or whatever) in a negative light, then you will have (inadvertently) directly trained your model to treat those features as negatives.

Of course, another (hacky) hypothetical (noted elsewhere in this thread) would be to use "resume + hire/pass" as your data set. In that case, your model would simply try to emulate your current hiring practices. If your current practices exhibit a notable bias towards a given feature, then your model presumably will too.