News for hackers. Stuff that matters.
Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by jacquesm on Wednesday December 02, 2015 @10:37AM (#10663544)
Is there anything about the false positive rate of this system?

Any binary 'good or bad' classifier should have at least a couple of stats attached to it to give an indication of how reliable the classifier is.

To complete the circle, it would be nice to know how it performs according to this:

https://en.wikipedia.org/wiki/Precision_and_recall

Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by kenrick95 on Wednesday December 02, 2015 @10:52AM (#10663643)
Not really, but one can report a false positives and false negatives here: https://meta.wikimedia.org/wiki/Research_talk:Revision_scori...

Since it is activated on Indonesian Wikipedia (I helped them to extend the tool to Indonesian language), I noticed that this tool can hardly capture an obvious vandalism yet. Some other edits are hard to verify whether it is vandalism or not, even by human. I believe this is a work in progress and hope that it will improve over time. But hey, it is the first time a vandalism detection tool is deployed in Indonesian Wikipedia. :)

Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by jacquesm on Wednesday December 02, 2015 @11:44AM (#10663994)
> Not really

That's worrisome then. You'd expect a bit more rigor before putting a bot like that into production.

Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by halfak_wmf on Wednesday December 02, 2015 @02:24PM (#10665187)
Actually, we openly report a wide variety of statistics about our models. E.g. if you go to https://ores.wmflabs.org/scores/enwiki, the service lists out the models that are available for English Wikipedia along with test statistics. You can also browse our test statistics in our documentation. See https://meta.wikimedia.org/wiki/ORES/damaging#English_Wikipe...

Currently, we're optimizing for AUC since it (1) is hard to cheat -- unlike accuracy or precision/recall individually and (2) it represents the ranking problem well.

See https://en.wikipedia.org/wiki/Receiver_operating_characteris...

"[AUC] is equal to the probability that a classifier will rank a randomly chosen positive instance higher than a randomly chosen negative one"

Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by jacquesm on Wednesday December 02, 2015 @02:33PM (#10665242)
Thank you. 75%, that's not bad but it's also not exactly 'x-ray eyes' grade yet. This is a pretty tough problem you're tackling, I can see now why it is not a bot but an augmentation device. It would be nice to track the accuracy over time, to see if the vandalizers wise up to how the bot makes the decisions to game it.
Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by halfak_wmf on Wednesday December 02, 2015 @02:56PM (#10665397)
Actually our AUC is 90% and that beats the state of the art in the literature. Accuracy is easy to game. We could get 97.5% accuracy by just predicting False all of the time. We don't advise reviewers to use the prediction directly, but rather to set thresholds on the probability. 70% probability is a very conservative filter and will catch about 10% and a little bit less than half of them will actually be damaging.

So, this is useful enough to reduce the workload by 90% right now. As we increase the model's fitness, we should be able to drive that up to 95%, but I doubt we'll get much better than that.

Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by jacquesm on Wednesday December 02, 2015 @03:10PM (#10665470)
> 70% probability is a very conservative filter and will catch about 10% and a little bit less than half of them will actually be damaging.

Right. This is really interesting. A while ago I wrote a classifier that predicted based on a number of criteria if a new member of a web community would be trouble or not. It also worked as an aid in classification and not as a stand-alone decision maker. You have a lot more data to work with so that should make the problem a bit more tractable.

Does it use other bits of knowledge about the edit prior probability of a revert for an edit by that particular user or network range for instance?

Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by tptacek on Wednesday December 02, 2015 @12:27PM (#10664338)
It's not actually doing reversions, right? It's just scoring edits. Wikipedia editors keep hot-lists of articles they watch edits for, or patrol lists of edits or page creations; ostensibly, all this needs to do is sort those lists of edits.
Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by jacquesm on Wednesday December 02, 2015 @02:38PM (#10665274)
It's augmentation, not making any decisions on its own (and that's logical given the false positive rate). So it's just an aid at this point but it can probably help quite a bit with triage.
Re:AI service gives Wikipedians ‘X-ray specs’ to see through... by kenrick95 on Wednesday December 02, 2015 @12:49PM (#10664515)
Yup, it's just scoring the revisions. Currently there is no straightforward way to feedback to the tool for learning false positives yet.