ABSTRACT Music mood classification supports recommendation, retrieval, and wellness applications, yet reported methods and evaluations vary widely across studies. Following the PRISMA systematic review methodology, this paper synthesises 39 peer‐reviewed works between 2009 and 2025 to summarise common practice across feature design, model families, datasets and reporting. Analysis reveals a fundamental structural split in the field: 17 papers focus on music‐centric classification (using audio, lyrics, and metadata), while 20 papers focus on user‐centric emotion recognition (using vision or physiological signals) to drive recommendation pipelines. Convolutional neural networks (CNNs) dominate the technical landscape (31 of 39), serving as a shared architecture for both audio spectrograms and facial analysis. However, this split creates a ‘mapping gap’ due to taxonomical inconsistencies between listener‐state datasets (e.g., FER2013) and music‐content datasets (e.g., DEAM). Reporting remains a significant bottleneck: 32 papers report only a single overall accuracy, masking uneven performance across mood classes, while only two papers provide granular metrics like F1‐score or recall. This study concludes that while technical pipelines are mature, the field requires standardised emotional mapping layers and richer evaluation metrics to improve the practical robustness and real‐world readiness of music mood classification systems.
Pitts et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: