Also, in training the system should tell you what it thinks you’ll think, and then you’ll correct it if it’s wrong. Without this you can overtrain, creating a continual positive feedback, where all you want is a corrective feedback.

OTOH, there’s rating systems, like (I think) on Amazon. There they correlate your ratings to other people’s ratings. There you can usefully rate anything, because determining your correlated users is separate from the rating it actually presents you — it’s not determining your preference, simply your demographic. Bloglines could do this (and they try a little), but you couldn’t in isolation. Anyway, some ideas.

For normal email spam/ham there are (or moreover, were) words that stood out as good/bad markers. But in something like an rss feed where you have already selected feeds that are reasonably close to your interest set, the individual words are no longer useful for sorting interesting vs yawn.

As you can probably tell, I had a go at doing this for an AI course I did last semester, and I wound up with a furball that didn’t actually fly. I used python and wxpy to build an aggregator around an IE control with good/bad buttons for marking, my word distributions didn’t wind up correlating.

Is your aggregator open for hacking? I have a few theories that I’d like to testbed, and now owning a mac, my old code is nigh on useless :-)

Intellie-Aggie tries to use these logics to filter out items, but I have left the codes untouched for far too long to expect any wonders.