Fortune Teller

In this paper, we document the development of a learning algorithm that rates the popularity of a song on a scale of 0 to 5, using a carefully chosen set of musical features. Along with personal interest, the main motivation is to be able to help both the artists and label/advertisement companies in the music industry. We considered two different datasets (MP3 and MSD) and we tried to mitigate the effect of unwanted factors as much as possible by carefully choosing the songs. We trained, tuned, and compared three different learning algorithms, and chose the SVM as it has the best performance. The results show that we can correctly classify a given song with 37.8% accuracy into the correct class, and in the event of misprediction, there is a 71% chance that the error is only one class away.