The role of artist and genre on music emotion recognition

Dissertation presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Knowledge Management and Business IntelligenceThe goal of this study is to classify a dataset of songs according to their emotion and to understand the impact that the artist and genre have on the accuracy of the classification model. This will help market players such as Spotify and Apple Music to retrieve useful songs in the right context. This analysis was performed by extracting audio and non-audio features from the DEAM dataset and classifying them. The correlation between artist, song genre and other audio features was also analyzed. Furthermore, the classification performance of different machine learning algorithms was evaluated and compared, e.g., Support Vector Machines (SVM), Decision Trees, Naive Bayes and K-Nearest Neighbors. We found that Support Vector Machines attained the highest performance when using either only Audio features or a combination of Audio Features and Genre. Namely, an F-measure of 0.46 and 0.45 was achieved, respectively. We concluded that the Artist variable was not impactful to the emotion of the songs. Therefore, by using Support Vector Machines with the combination of Audio and Genre variables, we analyzed the results and created a dashboard to visualize the incorrectly classified songs. This information helped to understand if these variables are useful to improve the emotion classification model developed and what were the relationships between them and other audio and non-audio features