Give
me personalized collaborative ranking. This is what I want:
resources of all kinds that are filtered and ranked according to people
I trust and respect.I assume it is a complicated problem, since I don't
have it yet, because I'm for sure not the first person to think about
it. But I believe it can be solved, if some capable person can work out
the math...[Ming the
Mechanic] Sometimes I think I'm on the same
Wavelength as Flemming. And yeah... sometimes that scares me a bit
;-) Anyway, again he's struck a chord that resonates in my own
source code tree something fierce. Those statistical algorithms for
matching sets of "interesting" items and weighting them according to
individual rating, then pushed and pulled based on both the "trust"
level and what Ming refers to as "qualitative judgements of people" who
are the sources of that information. (the two ARE orthogonal. For
example: I trust my sister implicitly, she wouldn't know a good movie
if it bit her in the ass.) Then there's the idea of recommending
across-media. If you and I have movie recommendations that are
precisely matched, should that promote our recommendations in other
media, or should that be configurable (well, maybe and yes). Enter
The Interest Engine. I've done alot of interface-free work over the
last several years in hammering these traits and behaviors into shape.
Most of the work I've done has to do with automatic classification of
text based on content analysis, but it's precisely the same thing. (In
essence the Engine "recommends" text bodies to categories based on lists
of key words and phrases common to other 'documents' in that
category.) What I've been doing is working with iTunes track
databases (you'd be surprised the kind of stuff people post on p2p
networks) and match music recommendations based on collaborative
filtering (hate that term.) So far my biggest problem isn't the code,
it's the fact that I can't find track & album listings that are close
enough to mine (or to each other) to produce meaningful
recommendations. The set math and statistics are just a bit off of
trivial once you've set up the abstraction layers right. An important
fact is that once those semantics are hammered out, there's no
particular binding on the type of information you're dealing with.
Bookmarks, songs, movies, books, restaurants, art, events, etc.
Everything's fair game. The problem? Data collection and
aggregation. It's not a technical issue at all. But it's a massive
buy-in problem. You've got to get people to recommend all these things,
to post their interests in some common format (say RSS and I'll smack
you. RDF maybe.) I'm perfectly willing to have an automated engine
posting my iTunes.xml file, a merge of my bookmark databases, etc.
But all of these other features Ming is talking about: The
configurable weighting of data sources based upon arbitrarily complex,
personally configurable criteria, is really no big deal. Anybody
have thoughts on the presentation layer? (ugghh, why does it always come
down to gui work :-/ )
1:18:38 PM 
|
