An Empirical Investigation of Stacking for Music Tag Annotation

Automatic tag annotation is one of the most important problems in multimedia information retrieval. It has been motivated by the large amount of unstructured tag annotation data provided by internet users and can be viewed as a variation of multi-label classification with special characteristics and constraints. Stacking is a technique in which the outputs (binary or probabilistic) of a set of binary classifiers (one for each tag) are used as input to a second stage of classification that attempts to exploit latent relationships between tags. This technique (known under a variety of names) has been used in a variety of multimedia tag annotation systems. In this paper we survey these approaches, clarify how stacking system are structured, and empirically investigate stacking using a variety of classifier combinations in the context of tagging pieces of music.