This is very cool - one quibble though, there doesn't seem to be any test of 'slow' in the slow build. You could add a linear regression slope and select for songs where Pearson coefficient is high, the slope is small and the last decile volume/first decile volume is high is large.