Collaboration-aware hit song analysis and prediction

• Autor(a) principal: Mariana de Oliveira Santos Silva
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/44585; https://orcid.org/0000-0003-0110-9924
• topic: Hit Song Science; Data Science; Machine Learning; Complex Networks; Music Data Mining; Computação – Teses; Ciência de dados – Teses; Aprendizado do Computador – Teses; Redes complexas – Teses; Mineração de dados (Computação) – Teses; Hit song science – Teses.

Hit songs are more successful than average, where key factors make such songs qualitatively superior to others. Current techniques to predict hit songs exploit features that describe songs individually. We propose tackling this prediction problem through a multimodal form with songs’ features fused, together. Specifically, we describe songs through features from three modalities: music, artist and album. Initially, we identify collaboration profiles in a musical network composed of successful artists, unveiling how artists professionally connect can significantly impact their success. Then, to deepen such analyses, we use time series and the Granger Causality test for assessing whether there is a causal relationship between collaboration profiles and artists’ popularity. Finally, we model the Hit Song Prediction problem as two distinct tasks: classification and placement. The former is a classical machine learning binary classification problem and is a direct application of our fusion strategies. The later is a modeling approach that ranks a song relative to a given chart, then predicts hit songs and provides comparative popularity information of a set of songs. Furthermore, we emphasize collaboration artists’ profiles as important features when describing their songs. Extensive empirical studies using various features from the modalities confirm the effectiveness of our method that fuses heterogeneous data for both tasks.