An Empirical Study of Multi-Label Classifiers for Music Tag Annotation

In this paper we study the problem of automatic music tag annotation. Treating tag annotation as a computational classification process, we attempt to explore the relationship between acoustic features and music tags. Toward this end, we conduct a series of empirical experiments to evaluate a set of multi-label classifiers and demonstrate which ones are more suitable for music tag annotation. Furthermore, we discuss various factors in the classification process, such as feature sets, frame sizes, etc. Experiments on two publicly available datasets show that the Calibrated Label Ranking (CLR) algorithm outperforms the other classifiers for a selection of evaluation measures.