Passive vs. Active Recommendation Systems
I had an email exchange with Chris Anderson a couple of days ago, author of the Long Tail blog (and of Wired Magazine fame) about recommendation systems. After sending Chris a pointer to my music recommendation system, he had this to say:
Thanks for the pointer. A quick bit of feedback: I’ve argued that the only recommendation services that will really scale are passive ones, that watch what people do rather than make them actively rate titles. Basically, I’m lazy and there are plenty of music recommendation services that don’t make me lift a finger. Is there any way you can make it zero-effort?
To which I replied:
I think you live too far in the future. One day all of our devices will play digital music and keep track of how many times we play song, songs we tend to skip over, your ratings for songs, etc. Until then, the only tools that really do that (well) are the iPod and iTunes.
Passive systems require you to already have your “profile” built, or at least “in the works.” A passive music recommendation system will work with all of your digital music players, collate everything they know about you, and then make recommendations. Today, I think very few people listen to music exclusively on their iPod or in iTunes. Even with the people who do use these players exclusively I still don’t believe their players have a very good idea of their listening tastes. You alluded to something like this in your “case against the shuffle” posting: what about the ghosts of “good” tastes past? The songs you forgot you liked? The songs that you hear on the radio, on TV, on CD, or at a friends house?
Until we live in a perfectly digital world I believe there will still be a need for active recommendations.
The problem with today’s passive recommender systems (I’m referring to MusicMobs and AudioScrobbler) is that, for most folks, they are totally worthless until you spend a month or two listening to all of your music on your iPod/iTunes. For most folks, it’s zero-effort => extremely slow gratification. Active recommenders are marginal effort => instant gratification. Have you played with MusicMobs/AudioScrobbler? You can’t get new recommendations until you’ve listened to at least ~50 more songs in iTunes. With GenieLab, you can sit there and go, “Well what if I rated Bob Marley a 3.5? Then what happens?” I think it’s a richer, more interactive experience.
I think Amazon is a perfect example of passive recommendations gone totally wrong. I mean, it’s good today–they’re light years ahead of everyone else–but their recommendations are based on an extremely limited view of your tastes. They only recommend books based on books you’ve bought or browsed at Amazon. (Plus they only recommend books that they have in stock, but that’s another debate)… What about all the classics sitting on your bookshelf that you’ve read a half dozen times? What about all those great books your friends loan you? Amazon has no idea who you are, they’re just firing shots in the dark. I don’t believe their “long tail” success has anything to do with their recommender system, I think it has a lot more to do with just the size of their online collection. There’s no cleverness in what Amazon is doing, it’s just volume.
I’m wondering if our little email exchange served as motivation for a recent post on Chris’ blog, where he writes:
…any service that tries to condense all of your different planes of influence into a single dimension is going to fail, at least as far as useful recommendations go. The filters that work best for me typically earned my trust by liking some of the same things I did, then turning me on to new stuff that I liked even more.
Chris is referring here to “friendship” networks that provide recommendations, and he’s arguing that they don’t work. Yet the recommendations he trusts are the ones that he sought out: he invested the time to discover who was recommending what, and decide if their tastes coincided with his. Isn’t this a sort of “active” system?
2 Replies to “”
this is a test comment…
FOAFING THE MUSICHere’s another music recommendation system, which merges both approaches: active and passive recommendations (well, actually, I’d rather say: “static and dynamic” recommendations).The idea is that the system recommends new music assets, based on what you like -static, it does not change “too much”-, and your listening habits -dynamic, it changes as you’re listening more and more music-.User profile and preferences are based on the FOAF project, whereas user’s listening habits are tracked from Audioscrobbler / Last.fm.