User recognition algorithm

Question

let's say you have a big IRC chan log, and you want to find out what user is using multiple accounts. As input you have the time the user connects to the server, and some sort of text analysis ( word frequency, and so on), and as output you want the likelihood two user "matches".

Is it possible to do it using ANN? Are there better algorithms to accomplish that task?

PS : use IP addresses is not an accepted solution :)

Answer 1

The problem with using neural networks is that you need a robust set of training data--that is, you need to have lots of examples of people using multiple accounts where you already know that's what they're doing. Furthermore, if the people you're trying to identify have ever played a role-playing game, they'll probably be able to make themselves seem quite a bit different if they want to.

So, if people are acting just like themselves and you have a pretty good training data set, then you stand a chance. You should probably start with methods used by forensic linguistics.

But I suspect that what you'll probably end up doing is identifying people who are sort of similar to each other. Good for a matchmaking site, perhaps; not so cool for most other things. (For example, I would think this would be a perfectly dreadful way to try to find members of Anonymous in other guises.)

Answer 2

This problem is known as "authorship detection" (or sometimes, in a particular domain, "plagiarism detection"). It can be done using a variety of statistical algorithms, of which neural networks aren't the easiest.

Check out the Cavnar & Trenkle algorithm for text classification. That may be made into a useful baseline algorithm for this task. Implementations in various languages are available on the web. You may want to turn it into a clustering algorithm instead of a classifier.