Presentation Information
[5M3-GS-2h-03]Music Emotion Recognition Models Based on Acoustic Features — A Study on Interpretability Analysis
〇YINGMING WANG1, Masayoshi Moteki1, Tengfei Shao1, Masayuki Goto1 (1. Waseda University)
Keywords:
Machine Learning,Music Emotion Recognition,Acoustic Features,Explainable Analysis
Music Emotion Recognition is an important research topic in music information retrieval and emotion analysis, aiming to automatically identify the emotional characteristics conveyed by music signals. Previous studies have improved emotion classification accuracy using machine learning models such as deep learning trained on large-scale datasets. However, despite these performance gains, the interpretability of the relationship between acoustic features and emotion recognition has not been sufficiently explored. From a practical perspective, understanding how acoustic features influence listeners’ emotional experiences is also important.
The objective of this study is to clarify the relationship between acoustic features and emotions perceived by music listeners. Acoustic features from the Million Song Dataset and user-assigned emotion tags from Last.fm are used to assign emotional labels to each track. Machine learning models are then constructed for emotion prediction. Using SHAP values, the contribution of acoustic features to emotion classification is analyzed, and their importance and directional effects are quantitatively evaluated.
The findings are expected to contribute to a deeper understanding of emotions in music information retrieval and music psychology, with potential applications in real-world contexts such as education and psychotherapy support.
The objective of this study is to clarify the relationship between acoustic features and emotions perceived by music listeners. Acoustic features from the Million Song Dataset and user-assigned emotion tags from Last.fm are used to assign emotional labels to each track. Machine learning models are then constructed for emotion prediction. Using SHAP values, the contribution of acoustic features to emotion classification is analyzed, and their importance and directional effects are quantitatively evaluated.
The findings are expected to contribute to a deeper understanding of emotions in music information retrieval and music psychology, with potential applications in real-world contexts such as education and psychotherapy support.