Regarding the gaydar paper, yes, I have read the full paper (if memory serves, I read two versions, a pre-print and the published paper). At the time, I wanted to publish a rebuttal, perhaps a letter in a journal or something, but in the end I didn't think I'd be adding much to the debate and the paper had been widely discredited already anyway.

My objection with the methodology in the paper was that the authors had assembled a dataset where the distribution of gay men and women was 50% of the population, i.e. there were as many gay women as straight and as many gay men as straight in the data. This was for one of their datasets, the one were everyone had a picture. There were two more where the distribution was less even but still nothing like what it's usually estimated to be. This despite the fact that the paper itself cited a result that gay men and women are around 7% of the population.

The reason for this discrepancy was clearly to improve the results by reducing the number of false negatives which are expected when there are many more negative than positive examples in binary classification.

This from the point of view of machine learning. There were other flaws that others pointed out, e.g. the choice of metric (I don't remember what it was now, I can look it up if you like), the premising of the paper on prenatal hormone theory that is another piece of bunkum without any evidence to back it etc.

And of course there were the ethical considerations.

Sorry but I don't have the courage to reply to the rest of your comment. You write way too much.

Rebalancing an imbalanced dataset is common in industry and academicia. You use that when you focus on accuracy, to make claims like: We were 54% accurate on classifying sexuality of females easily interpretable, without needing a distribution-balanced benchmark (you simply know it is a coin flip).

If there is signal in the rebalanced dataset, there should be signal in the imbalanced dataset. If they'd switched to logloss or AUC and an imbalanced dataset, do you think now their results would be as good as random? Because that is what you are implying and you are basically implying the research is fraudulent. This is a very strong claim to make, in the absence of legit discrediting studies that failed to replicate any predictability, and requires more than guessing the authors rebalancing act was "clearly" to improve the accuracy (with 7% negative class, you could get 93% accuracy by always predicting positive class, so if they wanted to inflate the accuracy, they shouldn't have rebalanced).

The ethical considerations are moot/personal opinion, as they passed the ethics board of Stanford. Those are people who evaluate ethics of academic research for a living, or are you saying they were also shoddy and wrong to give this a pass?

Magical thinking is not wanting something to be true, because it would be an uncomfortable truth, and so deeming that something which is objectively true, must be false, so you can continue to think happy thoughts in line with your world view.

You keep talking about the paper being widely discredited, but can't provide a single academic source for this. Instead, you question my sources (business insider?) while posting articles from The Next Web written by a History degree journalist who does not want the concept of binary sexuality to be true, or even allow it in constructing a dataset of gay and straight people by self-classification.

It takes more energy and letters to attack a point than to make a point. You made quite a lot of weak points.

>> Rebalancing an imbalanced dataset is common in industry and academicia. You use that when you focus on accuracy, to make claims like: We were 54% accurate on classifying sexuality of females easily interpretable, without needing a distribution-balanced benchmark (you simply know it is a coin flip).

You quoted The Strength of Weak Learnability and I figured you must have at least a passing acquaintance with computatinal learning theory. In computational learning theory (such as it is) it's a foundational assumption that the distribution from which training examples are drawn is the same as the true distribution of the data, otherwise there cannot be any guarantees that a learned approximation is a good approximation of the true distribution.

The following is a good article on machine learning with unbalanced classes:

http://www.svds.com/learning-imbalanced-classes/

I recommend it as a starting point.

>> This is a very strong claim to make, in the absence of legit discrediting studies that failed to replicate any predictability, and requires more than guessing the authors rebalancing act was "clearly" to improve the accuracy (with 7% negative class, you could get 93% accuracy by always predicting positive class, so if they wanted to inflate the accuracy, they shouldn't have rebalanced).

The gay class was the positive class and the straight class negative, in this case. If you did what you say and identified everyone as straight, you'd get a very high number of false negatives: you'd identify every gay man and woman as being straight. You'd get very high recall but abysmall precision. The authors validated their models using an AUC curve plotting precision against recall and such a plot would immediately show the weakness of an always-say-straight classifier.

>> You keep talking about the paper being widely discredited, but can't provide a single academic source for this.

An "academic source", like a publication in a peer-reviewed journal is not always necessary. For example, you won't find any peer-reviewed work debunking Yuri Geller. In this case my instinct is that no reputable scientist would want to get anywhere near that controversy (and that was one reason I also stayed away).

As to the work being widely discredited, the following is an article that summarises and links to criticisms:

https://greggormattson.com/2017/09/12/tracking-wang-and-kosi...

Some of the criticisms are technical, some are from the point of view of ethics. It would be a grave mistake to discount the ethical concerns, but if you prefer technical explanations there is quite a bit of meat there.

Thanks! That article has a lot of critique and I also like that the author collected the responses from one of the authors.

But, to me, most of the critiques seem uninformed (not made by ML practicioners) and focus on the ethics (where I agree with the authors: we need solid research into weaponized algorithms and show what is currently possible by ML practicioners, who may use such technology adversarialy, and can look at reclassifying profile pictures to the same degree as we do information about sexuality, religion, or political preference). By my estimation, most of the critiques are by people who find this research to be threatening to them, their friends, and their sexual identity. That may very well be the case, but it also leads people to conclude the scientific study was flawed and that an automated gaydar can't possibly work. Two replications by scientists who took issue with the paper, and lack incentive to fudge the data or metric to dress up their paper, also demonstrated a better than random automated gaydar. These systems work! (And that poses a problem we can now tackle, where before we did not even know this was possible, and the majority in this thread still thinks it is all bunkum).

Many statistical assumptions are regurarly broken, for pragmatic reasons (it just works better), or because the world is not static (and so the IID assumption is broken). There is an entire subfield of learning on imbalanced datasets, which includes resampling, subsampling, oversampling, and algorithms like SMOTE. It is common to use these techniques to get a better performance, including on unseen out-of-distribution data. Fraud - and CTR - and medical diagnosis models are regurarly rebalanced for other purposes than trying to break assumptions or cheat oneself into a seemingly higher accuracy. Plus, the signal does not dissapear when training only on originally balanced data. These systems do not work by the grace of a rebalancing trick alone, but they may work better (as usually the case with neural nets, which do not even give convergence guarantees: something only a statistician would worry about).

You can switch negative with positive class and my point remains: if the authors wanted the fraudulenty hack the accuracy score, this is way easier with imbalanced data. AUC metric robust to class imbalance anyway: ranking won't change for unseen data out of distribution, you can just adjust the threshold to match it.

I'd say an academic source is necessary in this case, because you implicitly accuse these scientists of doing shoddy hyped up work, with fudging tricks to appear more accurate. I need more than popular media sources or previous HN discussions to admit this paper was "widely discredited".

Your Yuri Geller example is a red herring: one is a stage magician, the other is peer-reviewed science. But to oblige: https://scholar.google.com/scholar?q="yuri+geller"

Yes, of course many theoretical assumptions are broken- but that is because people who break them either ignore them completely, or deliberately voilate them in order to produce better-looking results. That is more common in industry where it's easier to pull the wool over the eys of senior colleagues, but it's not unheard of in academia, quite the contrary. Anyway, just because people do shoddy work and then report impressive results doesn't mean that we should accept poor methodology as if it was good.

In particular about the gaydar paper, the authors cook up their data to get good results and then use those results to claim that they have found evidence for an actual natural phenomenon (hormones influencing haircuts etc). That's just ...pseudoscience.

Is your google scholar link humour?

You seem to be under the assumption that rebalancing is always bad or ignorant. That techniques, such as SMOTE, are only used to produce better-looking results and pull the wool over someones eyes. This is simply not true. Rebalancing is not shoddy, but accepted practice. It is certainly fair to question it, but not to draw the conclusion of fraud or shoddy science (without making you look pretty silly).

Again, I do not think rebalancing data justifies the conclusion that the authors were cooking up their data to report better results. Take a step back and assume good faith: could there be any other reasons to resample data, other than wanting to commit fraud?

The Google scholar links includes 10+ cited and peer-reviewed papers on the Yuri Geller drama.

I don't know enough about hormone theory to say anything against or for their conclusion, just focusing on showing that working automated gaydars that perform better than average/random guessing exist and have been scientifically demonstrated. I can agree with you on that the connection is spurious, without dropping my point that this controversial technology actually works (rebalanced or no).