Interpretability of Music Classification as a Criterion for Evolutionary Multi-objective Feature Selection

Abstract

The development of numerous audio signal characteristics led to an increase of classification performance for automatic categorisation of music audio recordings. Unfortunately, models built with such low-level descriptors lack of interpretability. Musicologists and listeners can not learn musically meaningful properties of genres, styles, composers, or personal preferences. On the other side, there are new algorithms for the mining of interpretable features from music data: instruments, moods and melodic properties, tags and meta data from the social web, etc. In this paper, we propose an approach how evolutionary multi-objective feature selection can be applied for a systematic maximisation of interpretability without a limitation to the usage of only interpretable features. We introduce a simple hypervolume based measure for the evaluation of trade-off between classification performance and interpretability and discuss how the results of our study may help to search for particularly relevant high-level descriptors in future.