Applying Active learning in Music Popularity Prediction

Abstract

Music popularity prediction has been widely used in the recommender system of various music platforms and is beneficial for artists to compose music. But the accuracy of the prediction is still inconsistent in previous research and most achieved low accuracy with a limited data set. This paper describes an approach for pursuing considerable accuracy with as few labeled instances as possible by using Active Learning. Starting with a data set from Spotify containing more than 6000 tracks and 15 features, the experiments in this paper firstly trained two different predictive models, and then use them to compare the learning progress of active learning algorithms with random selection. The results showed that active learning is beneficial for learning and improved the accuracy of the models. (Abstract)