Design and Evaluation of Semantic Mood Models for Music Recommendation using Editorial Tags

In this paper we present and evaluate two semantic music mood models relying on metadata extracted from over 180,000 production music tracks sourced from I Like Music (ILM)’s collection. We performed non-metric multidimensional scaling (MDS) analyses of mood stem dissimilarity matrices (1 to 13 dimensions) and devised five different mood tag summarisation methods to map tracks in the dimensional mood spaces. We then conducted a listening test to assess the ability of the proposed models to match tracks by mood in a recommendation task. The models were compared against a classic audio contentbased similarity model relying on Mel Frequency Cepstral Coefficients (MFCCs). The best performance (60% of correct match, on average) was yielded by coupling the fivedimensional MDS model with the term-frequency weighted tag centroid method to map tracks in the mood space.