Comment by quchen | Hacker News Reader

Comment by quchen | Hacker News Reader

Parent Full thread

I think the author's good intentions are taken a little bit too far here. As I understand it, it is claimed that the gender is "correct with 67% confidence", which I assume means "67% of all educated gender guesses are correct", with 50% standing for pure guesswork.

As is the case with any statistical data, a single value holds only very little information. The article does not do any analysis on how good these 67% are: could it be that changing the parameters changes it to 50%? There's a pretty large configuration space in question, and therefore I think giving out a simple number is very misleading.

What I also don't agree with is the conclusion, "Most importantly, if that's what I can do with a limited set of my own data, imagine what the NSA can do with the datasets it has access to." In the light of the previous paragraph, this draws the conclusion that the NSA can do pretty big things, but it's based on the wrong premise that the analysis of the "own data" has any statistical significance.

Note that I don't want to argue about the conclusion - data mining is very powerful, and so is statistics - but this kind of approach is very easily abused, be it intentionally or unintentionally.

So, in short: be careful trusting statistics, even when you really really like the result.

Another problem is he doesn't include any sort of meaningful baseline. Maybe 67% of the people he talks to are male. In which case, he's not predicting anything. Just guessing randomly biased by his own calling habits.

[EDIT: I had some other criticisms that were incorrect and overly harsh anyway. I've removed them.]

Thanks for the feedback. As I'm fairly new to all of this, I'd love to hear how I can turn this into a more viable experiment.

Also, to defend myself a little: I think I'm at least being responsible by in no way claiming that it's statistically valid, and in fact making the point that it's not, and citing the reasons why, several times in the article.

I do stand by my point, however, that this helps the public understand the danger of metadata for data-mining, as well as introducing them to the pitfalls of statistics.

> As I'm fairly new to all of this, I'd love to hear how I can turn this into a more viable experiment.

The most important thing is getting an estimate of how good your averages are. Try to modify a couple of things, for example how do the 67% change with sample size (verification)? How does the number turn out if you feed it biased data, e.g. only male phone numbers (falsification)?

Thanks. I solemnly swear to not abuse statistics, so I'll play around and update the piece.

The goal is to use this as a hook to get the public interested, then take them for the ride as I learn. That's why I tried to be really cautious about pointing out all the reasons why this particular model is bad.

I didn't mean to say you specifically were abusing statistics, it's just very easy to draw dubious conclusions. (Doing this intentionally would be the abuse I was talking about.)

I know, but I still want to take it seriously, and thank you for your criticism.

Thanks to your help, I found a huge, obvious problem that was keeping the accuracy low and remarkable stable no matter what size of training data I tried. Here's the update: http://www.fastcolabs.com/3012908/tracking/im-beating-the-ns...

Reply on news.ycombinator.com