A system for automatically generating music playlists based on a user selecting desirable or undesirable music.
- Inferring Similarity Between Music Objects with Application to Playlist Generation by R. Ragno, C.J.C. Burges, C. Herley, Proc. ACM workshop Multimedia Information Retrieval, pp. 73-80, (2005).
- Fast Embedding of Sparse Music Similarity Graphs by J. C. Platt, NIPS 16, pp. 571-578, (2004).
- Learning a Gaussian Process Prior for Automatically Generating Music Playlists by J C. Platt, C.J.C. Burges, S. Swenson, C. Weare, A. Zheng, Advances in Neural Information Processing Systems 14, pp. 1425-1432, (2002).
The AutoDJ project created a system that helps you enjoy your personal music collection. AutoDJ takes one or more seed songs, and generates playlists out of your own music that incorporate the seed songs, and chooses other songs that fit well with the seed songs.
Notice that playlist generation is not the same as music recommendation: playlist music is a disconnected scenario. We need to generate a playlist out of your own music on a standalone player, with no access to the Internet. This is a tougher problem, for two reasons: 1) we have to be careful of gaps in your music collection, because they can create odd playlists, and 2) all information to create the playlist must reside on a player, without immediate access to a server.
An obvious approach to solve problem (2) is to base playlist generation on audio analysis of music. However, when we started the project, audio analysis of music was far below the accuracy bar to generate good playlists. (Audio analysis gets better each year, see the Music Information Retrieval Exchange web site for comparative results.)
Instead, we decided to analyze all music metadata that was available to us, to find music that is similar to each other. In our 2002 paper, we started with music that was hand-labeled with attributes. Our 2003 paper extended this to a graph of similar music, which gets compressed and sent to the player once. The 2005 paper analyzes