Comparing Fuzzy Rule Based Approaches for Music Genre Classification

Abstract

Most of the studies on music genre classification are focused on classification quality only. However, listeners and musicologists would favor comprehensible models, which describe semantic properties of genres like instrument or chord statistics, instead of complex black-box transforms of signal features either manually engineered or learned by neural networks. Fuzzy rules – until now not a widely applied method in music classification – offer the advantage of understandability for end users, in particular in combination with carefully designed semantic features. In this work, we tune and compare three approaches which operate on fuzzy rules: a complete search of primitive rules, an evolutionary approach, and fuzzy pattern trees. Additionally, we include random forest classifier as a baseline. The experiments were conducted on an artist-filtered subset of the 1517-Artists database, for which 245 semantic properties describing instruments, moods, singing style, melody, harmony, influence on listener, and effects were extracted to train the classification models.