Open jgomezm opened 4 years ago
Will track level features like danceability be used? https://developer.spotify.com/documentation/web-api/reference/tracks/get-audio-features/
I think that they can be added as additional features outside of the LSTM cell, but am not sure how to do it. I also noticed that we're keeping loudness which has 3 measures and confidence which is a feature of the algorithm, not the song
Edit: Removed loudness and confidence measures
For confidence, I thought that our model can develop an understanding of how much to trust the provided data point. If the algorithm that generated those features is not confident about an observation, LSTM may choose to undermine that information.
What is wrong with loudness?
I don't like confidence only because it isn't a feature of the music but of the algorithm. If their algorithm has a bias that impacts its confidence, then our algorithm can learn to exploit that weakness. In which case, it won't learn to discriminate between tracks if it is easier to discriminate between segments of confidence. Similar to the mean pixel value example in the DL PS.
We can keep one measure of loudness, but I think that 3 is too much. Wouldn't know which to keep. Any preferences?
Maybe loudness_max?
Pull data from Spotify and structure accordingly https://developer.spotify.com/documentation/web-api/reference/playlists/