What's the best logic model to use for a recommendations engine? Builders of popular music recommendations services shared the pros and cons of their own - and each other's - approaches to matching people to the music that they'll love. These four paradigms for recommendations engines came out of the conversation. Most recommendation services will fall into one of these models, but the 'best approach' would undoubtedly vary based upon the subject matter.ight expect, MusicIP licenses its technology for pushbutton playlists on many consumer devices.
#1: Trust your friends
Ali Partovi started iLike believing that it was important to have both computer- and friend-generated recommendations. Over time, however, it has turned out that users find the friend recommendations to be much more compelling. There's plenty of ways to improve the math of this, but the basic paradigm is that your recommendations are derived from both the music your friends have recently played, and what tastes you have in common.
#2: Heed the ivory tower
Under the 'Musical Genome Project,' the experts at Pandora categorize music using over 600 taggable properties. They collect songs to tag based on direct submissions, scouring top charts, and reviewing what new music their users have collected. When you enter an artist's name that you like, Pandora starts playing music that the experts think you'll enjoy based on tag matching. The system can also learn, and the Pandora player accepts your thumbs up/thumbs down on selections.
Pandora does play an editorial role. As Tom Conrad notes, 'just because you might like Chicago, it doesn't mean that you want to hear every song that Chicago every played.' There's also a sense that an editorial hand can steer away from the 'Radiohead problem': how do you make sure that all recommendations don't lead to Radiohead, if Radiohead is popular with a high number of users?
#3: Spy on the community
Last.fm has a huge community of music fans. Since the service is focused on capturing a complete musical profile on each listener, including everything that you listen to, it provides crowd wisdom on what's worth listening to. Like following the news on digg, last.fm lets you be a voyeur into what music the community is following. Some may consider this discovery process to be a bit random, since community tastes aren't necessarily what you would expect, but it does provide a refreshing level of discovery. Because of this, Martin Stiksel hopes that last.fm will 'be the last music platform that you'll ever need.'
#4: Follow the math
MusicIP offers recommendations based on core elements of the music composition itself. This helps you explore your own music library better by understanding the relationships between songs - if you're in a certain musical mood, you won't get stuck listening to the same album over and over. As you m
