The efficacy of familiarity as a feature in a music emotion recognition (MER) system was evaluated by a Random Forest feature importance analysis on a novel dataset of 5000 clips with annotated familiarity and valence. Familiarity was correlated to perceived valence (r=0.250) and resulted in a statistically significant increase of 0.011 in the F-score of a baseline MER classifier upon its inclusion. This work was presented at the Music and Machine Learning session at the European Conference of Machine Learning, Wurzburg Germany, 16 Sept 2019. Full paper available here.
Fig. 1. Feature importance rankings for all 58 features