Further musings toward smarter aggregators
I'm a complete neophyte when it comes to machine learning, but I'd like to get into learning more about the field in general. In particular, I'd like to make my news aggregator smarter. I've already tried using SpamBayes, but that didn't make me happy. Whether it was my approach or whether it was that Bayes itself is not suited toward this task, I'm not sure, though I suspect it's a little of both.
It seems like the magic Bayesian pixie dust works well for spam-vs-ham in my email box, so why shouldn't the magic for interesting-vs-yawn work for my aggregator firehose? Well, here are the issues I'm guessing at:
In the case of spam-vs-ham, you want to classify things into this or that-- that which is kept, and that which is tossed away. But in the case of items in my aggregator, I want a relative sort order or a score. I want a fuzzy guess toward my interest with which to inform presentation of items. Interesting-vs-yawn is more of a continuum than a pair of buckets.
And then, there's the passive gathering of behavioral data from my interactions with the aggregator, because I'm sure as hell not going to click ratings or thumbs-up/down all day. In spam-vs-ham, I could build up two clean mailboxes for training the categorizer, with one containing all spam and the other all ham. But, in the case of my aggregator, the only thing I'm tracking are items in which I showed interest by revealing more information or by clicking through.
So, I can say that a particular pile of items are all interesting. But, my interest level for the rest of the items received is a complete unknown-- maybe I'm vehemently disinterested in those 50 items, but maybe I just never got around to looking at those other 20 and just let them fall off my date range for display. Thus, I have a pile of ham, and a pile of undifferentiated unknown. I'm not bothering to provide any cues as to whether I don't like something, because that'd be boring work-- I mean, I am disinterested in those items, after all. So, I'd like to leverage what the system knows from what I care to provide, but not jump to any conclusions about the items in the unknown pile. There is no spam, only various flavors of ham.
Given all this, then, is there anyone out there who knows more about machine learning than me who could maybe point me toward a better approach or algorithm that fits this profile?
shortname=further_smart_aggregator_musings
Archived Comments
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.