Too bad there are no details about the algorithm. I'd use something supervised, for how it's described in the first part it looks some sort of bayes