This study investigates the extent to which machine learning models can predict tag-derived emotion categories in music using only Spotify-derived acoustic features and how those models rely on different acoustic cues when making predictions. Leveraging Last.fm user- generated tags, a transformer-based classifier (DistilRoBERTa) was applied to generate weak emotion labels covering six basic emotions: joy, sadness, anger, disgust, fear, and surprise. After Spotify matching, the analytic dataset includes 82,950 emotion-tagged records and 39,635 unique tracks with one or more emotion labels. A multilabel top-k evaluation strategy was used to reflect the non-exclusive structure of musical emotion labels. Three models, Random Forest, K-Nearest Neighbors, and Multi-Layer Perceptron, were trained and compared. Random Forest achieved the strongest overall performance, with higher F1 scores for frequent labels such as joy and sadness, while lower-frequency and semantically ambiguous categories, especially fear and surprise, remained more difficult to classify. Feature importance, SHAP-based interpretation, correlation analysis, and ablation testing showed that acousticness, energy, loudness, and valence contained much of the predictive signal for the weak labels, while errors often arose from overlapping labels or conflicting acoustic cues rather than random noise. The study is limited by its reliance on transformer-generated labels, exclusion of lyrical and higher-level structural features, and possible genre-specific tagging biases. However, these constraints also provide a useful test of how far weakly supervised, audio-only models can go when applied to large-scale listener generated music data. Overall, the findings support the feasibility of weakly supervised, audio-based multilabel emotion classification and provide interpretable insight into how machine learning models associate acoustic features with tag-derived emotion categories.
Jiaming Mao (Mon,) studied this question.