How to get it?

• There were some public challenges that released huge gobs of data: link prediction task.
  • They include some interesting song features that encode some "deep" content of what the song is, for neural net training
  • Here is an example analysis
  • All the analyses submitted for this challenge are (by requirement) open source
• Top 50 by year
• Hit predictor dataset has 40k songs labeled "hit" or "flop"
• All the songs has 170k+ songs

What to do with it?

• First, need to isolate a behavior (so we're not just analysing the acoustic shapes of albums, etc.)
  • This includes a skip, or starting on a specific song (i.e. the beginning of a session)
• Second, there should be some interesting datapoint
  • Maybe this just has to do with who listens to what? Cultural grouping etc.