Avaliação de características e previsão de sucesso de canções populares brasileiras por meio de aprendizado de máquina

• Autor(a) principal: Bertoni, André Augusto
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFG
• dARK ID: ark:/38995/00130000091cd
• Texto Completo: http://repositorio.bc.ufg.br/tede/handle/tede/11387
• topic: Ciência das canções de sucesso; Aprendizado de máquina; Extração de características; Recuperação de Informações Musicais (RIM); Melodia predominante da voz; Hit song science; Machine learning; Feature extraction; Music Information Retrieval (MIR); Predominant voice melody; ENGENHARIAS

This work aims to develop algorithms capable of predicting whether a Brazilian popular song can become a commercial success, using the help of Machine Learning techniques. To achieve this goal, a bank of popular Brazilian songs performed on the radio from 2014 to 2019 was created, applying the number of radio plays as the criterion for separating the songs into a successful group and another of non-successful. Several techniques of Data Analysis were studied and applied to optimize the databases and extract statistical characteristics of the songs. From the study of music theory, a set of musical semantic characteristics extracted from each song was also defined to support the Machine Learning algorithms, and then employ Data Science techniques to predict if a song can become a commercial success. Classification algorithms with supervised training were used, both by the classical approach and by means of Deep Neural Networks. For training and validation, cross-validation was used with ten subsets for the classical approach, and five subsets for convolutional networks. The performance of the algorithms was compared basically in terms of accuracy, precision, sensitivity and specificity. The discussion of the results of this work showed that statistical characteristics extracted from the songs brought satisfactory results in several metrics, such as: Accuracy (69.63%), Precision: (69.03%), Sensitivity (71.55%), Balanced Accuracy (69.75%) and ROC (69.75%), using classic techniques as: Naive Bayes, Decision Tree, Random Forest, kNN, Logistic Regression, SVM and MLP - which represents an excellent result, when compared to several other works of literature. Deep Neural Networks of the Convolutional type did not bring good results, with little better accuracy than randomness. The best scenario was achieved by combining three distinct banks of characteristics: a) statistics; b) spectrographs extracted from the Main Voice Melody; c) Musical Melodic Semantic information. With the combination of these three distinct banks of characteristics, 74.54% Accuracy was obtained.