Multi-label Learning Approaches for Music Instrument Recognition

摘要

This paper presents the two winning approaches that we developed for the instrument recognition track of the ISMIS 2011 contest on Music Information. The solution that ranked first was based on the Binary Relevance approach and built a separate model for each instrument on a selected subset of the available training data. Moreover, a new ranking approach was utilized to produce an ordering of the instruments according to their degree of relevance to a given track. The solution that ranked second was based on the idea of constraining the number of pairs that were being predicted. It applied a transformation to the original dataset and utilized a variety of post-processing filters based on domain knowledge and exploratory analysis of the evaluation set. Both solutions were developed using the Mulan open-source software for multi-label learning.