Unsupervised Heterogeneous Graph Neural Network for Hit Song Prediction through One Class Learning
Angelo Cesar Mendes da Silva, Marcos Paulo Silva Gôlo, Ricardo Marcondes Marcacini · 2022
Although the concept of success is subjective, it can be related to the popularity and interest of users. Measuring the success of a song in advance allows for offering information of great interest to the music market. Hit song prediction is an existing task in Music Information Retrieval that explores approaches for measuring music success based on features. Musical data is intrinsically multimodal, where features from different sources have complementary semantic information. Therefore, structuring musical data and building a unique space that embeds multiple features is a challenge in musical data representation. Using heterogeneous graphs to structure multimodal data is a resource for exploring the intrinsic semantic relationship between features. In this sense, this work proposes to structure musical features over heterogeneous graphs and learn a new graph-based multimodal representation for songs using an unsupervised graph neural network to handle the hit song prediction task. We formulated the hit song prediction task as a one-class learning problem to mitigate the non-hit song gaps and highlight the hit song as the interest class. We measure the performance of representations based on lyrics and artist features and present promising results using our learned representations that outperform other strategies for representing musical data.