Mapping Musical Mood with Unsupervised Learning: PCA Spaces and Cosine-Similarity Recommendations

We ask whether lightweight, explainable methods can model musical mood and support mood-aware retrieval without raw-audio pipelines or listener ratings. Using the Coimbra MIR 4Q dataset annotated under Russell’s circumplex (900 clips), we merged musically interpretable features that summarize tempo, timbre, rhythm, and dynamics, alongside compact tag encodings. After standardizing feature blocks, we applied principal component analysis (PCA) to obtain a low-dimensional embedding; loadings suggested that PCA1 tracked dynamics and meter steadiness, and PCA2 tracked rhythmic variability. Without using labels to fit PCA, the two-dimensional map aligned with the four circumplex quadrants. Model selection used scree and reconstruction-error curves, which indicated diminishing returns after about 6–8 components. Treating quadrants as clusters for evaluation yielded strong separation (silhouette 0.609, Davies–Bouldin 0.483, Calinski–Harabasz 3661.9). A cosine-similarity recommender retrieved nearest neighbors that were musically and emotionally coherent, with an option to emphasize items near quadrant boundaries to surface blended emotions. Because the approach remains in a tabular feature space, it is transparent, fast, and easy to tune through feature weights and tag contributions. The results demonstrate a practical path to mood-aware recommendations using explainable techniques and publicly available features.