Filtering:

1. use only first track of each release
2. use only one track per artist
3. use everything we have

Balancing:

1. balance dataset, to equal number of track per class (random subselect)
2. do not balance dataset, but preserve genre ratios across all acousticbrainz collection

Using sub-genres

1. train on lots of classes (for examples, 50 subgenres) to see which classes are frequently confused (mis-classified): probably these classes cannot be separated acoustically using our current features.

Genre ground truth source

1. use musicbrainz recording tags
2. use genre information from other sources (last.fm)
3. use genre clusters that are acoustically separatable according to our sub-genre classifier
4. ask people to create genre clusters and assign recordings to each genre