A machine learning approach to musical success prediction and genre emergence

In the past years, Spotify and YouTube have been dominating the music production scene, thus storing valuable information about how songs achieve popularity and become a successful hit. Past researches tried to determine the hit-potentiality focusing on the audio characteristics and showed the limitation of this approach. In this work, we studied popularity dynamics of songs including, together with audio characteristics, features from artist and release context, and processed these using machine learning techniques. As metrics for popularity, we employed the cumulative views on YouTube achieved by a song's video in this streaming platform. First, we investigated the relationship between songs with similar audio feature and their release date to identify sub-genres with concomitant popularity. Then, we processed our data by fitting the songs' popularity, ranked by their followers' number on Spotify, using a Zipf's model to provide a reference for a song's success accounting for the fame previously achieved by its artist. Thus, comparing the actual performance against the one predicted by the Zipf's law, we defined a success score whose forecast, using different machine learning models, could determine which features had most predictive potential. From our results, release dates showed little to no relation with music sub-genres, as the distribution of release dates by sub-genre peaked in almost the same times. Namely, songs are released irrespective of their genre. Moreover, while the correlation between YouTube views and followers' number on Spotify was moderate, stronger dependence resulted from songs with negative Success score. We hypothesised that these songs only achieve ``business as usual'' performance, as opposed to those with positive score that outperform instead. Using supervised learning techniques, we showed that audio and contextual features from Spotify and YouTube (e.g., the number of subscribers of the YouTube channel featuring the song) were sufficient to train a model capable of determining with good accuracy the Success score of different songs: in particular, a great influence on the prediction was determined by the YouTube context in which the song was released.