I hope the researchers a re well-meaning and that the study is more sophisticated than the abstract makes it seem, but it seems to have multiple potential problem areas:
One-dimensional "extremism." Can they detect "white power" recruitment? Environmental extremist recruitment? It seems too easy to rate "angry and in the Middle East" as "terrorist."
If it's too easy to score a hit, it may be too easy to avoid a false positive. What's the system's real capability of avoiding a false positive?
There is no comparison to current "best practices" such as they are. How does the FBI select extremists for surveillance? Does this do a better job?
You can see the curve of false positives and false negatives in the paper. As one example, finding half of the extremist recruiters involves about labeling about 2% of benign posts as extremist recruiters. It's not promising if you think that they'll automatically suspect people of being terrorists based on this classifier, but it could be used responsibly as a filter for a human to follow up on.
I assess from the paper that they're specifically looking for Islamist extremist recruiters and not even considering any other type of extremists in this model.
Overfitting to suspect people with Arabic names would be a concern. (You know the classic 20 Newsgroups classification problem? Any algorithm that does particularly well at it is just overfitting to the names and signatures of people who posted in each newsgroup.)
2% false positives? That's terrible, because it's not scaled to the ratio of benign posts to recruiting posts. That is, at a 2% rate, false positives could, perhaps vastly, outnumber accurate hits.
One reason I mentioned other kinds of extremists is their method of using judges to manually rate posts. It could well be that Americans, in particular, have a media-formed consensus on what an islamist extremist is. Tree-sitters, antivivisectionists, anti-abortion, etc. "extremism" could be harder to judge, and that could jack up the error rate considerably.
This currently goes to a landing page that makes it look like you need to sign up first to read the article. That's not actually true, clicking 'full text' on the right hand sidebar will take you to what may have been a better link to post here:
There is no requirement to sign up. The link is to the abstract, but if you read it and find it interesting, there are two links to the full paper, as both pdf and text.
One-dimensional "extremism." Can they detect "white power" recruitment? Environmental extremist recruitment? It seems too easy to rate "angry and in the Middle East" as "terrorist."
If it's too easy to score a hit, it may be too easy to avoid a false positive. What's the system's real capability of avoiding a false positive?
There is no comparison to current "best practices" such as they are. How does the FBI select extremists for surveillance? Does this do a better job?