Archive of UserLand's first discussion group, started October 5, 1998.
Re: Collaborative filtering and Napster and Firefly and ting
Author: Yoz Grahame Posted: 7/14/2000; 6:44:53 AM Topic: Collaborative filtering Msg #: 18602 (In response to 18541) Prev/Next: 18601 / 18603 I think Napster's got a wonderful unintelligent collab-filter: not only from looking around people's collections when I see they've got interesting rare tracks, but from watching who's downloading from me and what they're grabbing. If I see someone downloading Tortoise, 808 State and Lee Perry from me I know they're going to have a similar taste in music and so I go and look at them. (Although it'd be a hell of a lot easier if "sort by username" worked on the download screen)
I'm still not entirely convinced that collab-bots like Firefly can't work... I just think that they weren't very well tuned. They'd almost always give you stuff that was quite good but relatively middle-of-the-road, and no use for hunting down things in particular genres because genres are so hard to do well in music these days. And also because the users didn't have much control over their settings, and they could only rate on a single linear scale.
The frustration I had with Firefly was the usual "Yeah, I think REM's good, but I think Bark Psychosis is brilliant" problem. Collab filters drive you away from niches, towards the middle ground, which is pointless in a recommendation system, at least it is for me, because I'd rather be recommended interesting niche stuff most of the time, but instead it gives me Elvis Costello. Or maybe it's being a kind of classics teacher, telling me I should spend some time with the greats before I wander off looking for quick thrills. (This is where my guilt for not listening to enough Beatles/Stones/Curtis/Public Enemy comes in)
Here's an idea, with the usual discalimer because it's only just come to me and I'm really tired: how about a Firefly where you can make up your own rating scales for music, as many as you like - so you could, if you wanted to, rate music on overall quality, vocal performance, cheesiness, technoness, chilledness, etc. You, as the user, make up whichever scales you like for music and then you rate stuff on those scales. And you don't have to explain what these scales mean, at least, not to the bot - you can put in a description string so that other users browsing your scales can see what they mean, according to you. And then, um, something like having the bot compare scales between people and find matches in the scales, rather than matches on individual albums - so if I have a scale called "technoness" and you have a scale called "bounciness" and the bot compares our ratings on those scales for albums and sees that we've both given Basement Jaxx a high score and Johnny Mathis a low score (except with a much larger sample base, obviously, to get some differentiation between "bounciness" and "grandma-wouldn't-like-it-ness") and then decides that the two scales are roughly equivalent. Then you can log on and say "I want a new album, something good, quite bouncy, with occasional chilled bits", except you say that by doing it in terms of the scales you've defined, and it goes out and looks for scales that match your scales.
It'd be interesting to build the sort of arbitrary-scale system I talk about into Napster, with no bot intelligence at all - just a way you can show people browsing your stuff how you rate each track. (That's the other problem with rating albums rather than tracks - saying "occasional chilled bits" as opposed to "the whole thing is averagely chilled" is impossible with a single chilledness score rating an entire album)
-- Yoz
This page was archived on 6/13/2001; 4:55:40 PM.
© Copyright 1998-2001 UserLand Software, Inc.