Novel Datasets for Evaluating Song Popularity Prediction Tasks

We present two novel datasets for hit song prediction (HSP): HSP-S and HSP-L. They are substantially larger than currently available dataset with respect to the number of features contained. Both datasets provide high- and low-level audio features stemming from AcousticBrainz and short representative MP3 samples. Further, we include listener- and play-counts gathered from last.fm for both datasets. The larger dataset, HSP-L, contains 73,482 songs with audio features, listener- and play-counts, making it substantially larger than previous datasets. In addition, we provide the release year information for 65,575 songs as provided by the Million Song Dataset.<br> The smaller dataset, HSP-S, contains 7,736 songs, listener- and play-counts as well as Billboard Hot 100 data for 50% of the songs, and release year information is available for 7,449 songs.