Análise multivariada do perfil musical com dados do Spotify

Abstract

This study analyzes musical characteristics and their relationship with the popularity of songs over several decades, using multivariate statistical techniques, supervised machine learning (Linear Regression), and unsupervised machine learning (PCA and Clustering). Variables such as valence, energy, danceability, instrumentality, and others were analyzed to understand how they influence musical success. Through Principal Component Analysis (PCA), it was identified that the main dimensions of the dataset are associated with attributes such as energy, danceability, acoustics, and instrumentality. The correlation matrix revealed significant relationships between variables, such as the strong correlation between energy and loudness and the negative relationship between instrumentality and popularity. Additionally, Factor Analysis was applied to identify latent dimensions that explain the interrelationships between musical variables. The extracted latent factors highlighted groupings of characteristics, such as emotional positivity and sound dynamics, providing insights into the underlying patterns that influence song popularity. Furthermore, cluster analysis was also employed using the non-hierarchical K-means method, resulting in five groups with similar characteristics, ranging from introspective and acoustic songs to highly danceable and energetic ones. The evolution of musical popularity over the decades was also analyzed, highlighting changes in attributes preferred by audiences. The results provide valuable insights for the music industry, emphasizing consumption patterns and factors that influence the success of songs. This study contributes to understanding the dynamics of popular music and proposes a methodological foundation for future research. As limitations, the absence of external variables, such as media impact and cultural context, is noted. Future studies may integrate qualitative data and external variables to expand the understanding of the analyzed phenomena.