An unsupervised feature selection approach for finding diverse emotional-semantic representations in sonic branding music

Discovering the emotional–semantic dimensions underlying music description is central to music psychology and widely applied in sonic branding practice. Academic work typically relies on dimension reduction approaches such as principal components analysis ( PCA ), which (i) require subjective reinterpretation of latent components, (ii) often yield uneven component importances when a fixed number of dimensions is imposed, and (iii) offer limited guidance for selecting practically manageable subsets of descriptors. Addressing this gap, we evaluate whether an existing feature-selection algorithm — Diversity-Induced Self-Representation ( D-ISR ; Liu et al. [2017]) — can serve as an objective and scalable alternative for identifying concise yet representative sets of emotional–semantic attributes. Using a large real-world dataset (N Participants = 55,593; N Responses = 5,820,188; N AudioTracks = 251), we compare D-ISR and PCA within a unified experimental framework. D-ISR selects 14 core attributes from an industry-scale pool of 212 attributes and reconstructs the original 212-dimensional space with good accuracy. Direct comparison with PCA demonstrates how D-ISR provides a more balanced trade-off between interpretability, reconstruction fidelity, and the need for a practically small set of descriptors. Our findings (i) document and analyse a large-scale emotional–semantic music dataset rarely accessible in the public domain, (ii) demonstrate a principled framework for comparing feature-selection and component-extraction methods for music-descriptor research, and (iii) illustrate how large attribute sets can be reduced to flexible, task-appropriate representations of the emotional–semantic music space. This contributes a clear methodological foundation for both scientific studies of musical meaning and applied work such as sonic branding.