Machine learning methods in music emotion recognition

Abstract

Music Emotion Recognition (MER), an area within Musical Information Retrieval (MIR), studies the emotions evoked in listeners by music. We address MER as a regression task, with the objective of predicting the emotional content of music (encoded in arousal and valence) from acoustic features extracted from the waveform. We apply an interpretable machine learning technique, investigating the role of these features in predicting the target variables. Initially, a random forest model is trained on the DEAM dataset (MediaEval Database for Emotional Analysis of Music). Then, we use the concept of Shapley values to interpret the role of each variable in the predictions made by this model. Finally, we extract the most significant features from the DEAM dataset to predict arousal and valence, thus enhancing the interpretability of the model employed. Additionally, we explore a dynamic linear model approach to gain further insights into the relationships between features and response variables. This method allows for a potentially “less black-box” and more interpretable representation of the problem. Principal Component Analysis (PCA) is also utilized to analyze the structure of features in the dataset, providing a more comprehensive understanding of the key variables influencing MER predictions. By integrating these approaches, we aim to enhance both the predictive performance and interpretability of the models, offering meaningful insights into the most relevant features that drive emotional responses in music.