- output an infinite sequence of track choices that is enjoyable to listen to, without wanting to change any settings
- do this by learning, based on what the user does
- dont think of this as a permutation of all songs. it is acceptable and desirable not to play horrible songs at all.
- letting an entire track play is a vote in favour. skipping to the next track is a vote against.
- the pool of tracks is fixed, no new tracks are added
- could generalise later
- ignore track rating metadata
- initially treat all tracks as independent. ie no recommendation model to generalise between tracks
- could generalise later with contextual bandit approach
- per-albumn / per-artist features?
- assume song preferences are invariant wrt mood/etc of listener
- could generalise later. probably need contextual bandit approach
- seed playback sequence with initial manual track choice(s)
- try to infer a (relative?) desired playback frequency for each track
- maintain playback history to measure frequency each track has been played
- maintain votes for and against tracks (by listener actions / inactions)
- use epsilon-greedy or UCB approach
- use some kind of structured prediction approach? SEARN?? to generate sequences of tracks
- foobar2000 (windows, really nice, has SDK for plugins, C++)
- quodlibet (cross-platform, GTK, GPL, supports plugins, python)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent