Consultation des notices

We present a system that can learn effective classification models from music databases of very different characteristics, including both single-label collections indexed by genre or artist and multilabel databases of musical mood and instrumentation, where multiple tags can be applied to each track. Adaptability is attained by means of automatic feature and model selection, both embedded in the multiple-instance binary relevance learning of a Support Vector Machine. We discuss strategies for compensating overfitting and unbalanced training sets.