Hit Song Science Is Not yet a Science

We describe a large-scale and complete experiment aiming at validating the hypothesis that the popularity of music titles can be learnt using global acoustic or human features. We use a 32.000 title database with 632 manually-entered attributes per titles including 3 related to the popularity of the title. We design an experiment with two different audio feature sets, as well as the set of all the manually-entered attributes but the popularity ones. The experiments show clearly that although some subjective attributes may indeed be reasonably well learned by these techniques, it is not the case for popularity. This contradicts recent and sustained claims made in the MIR community as well as in the media about the existence of a so-called “Hit Song Science”.