After doing some experimentation with Bayesian classification of HN articles (for selecting articles that interest me) I've actually found that titles, rather than URLs, domains, or even article text, is one of the best indications of how good an article (and the accompanying discussion) will be.
(This is one of the reasons that I hate the moderators reverting article titles to less descriptive ones.)
It's a little bit of both submission title and domain. In this particular case, the title contained an "I": when people looked who the "I" was, they immediately upvoted.
I haven't, but I can if you're interested. "Best results" in this case is pretty subjective so I didn't perform any rigorous experiments about the relative quality of each method, that was just a general observation that I had.
The program itself is just a quick HN scrapper combined with a naive bayes classifier which periodically runs and sends selected article titles to an IRC channel.
(This is one of the reasons that I hate the moderators reverting article titles to less descriptive ones.)