Melodic track identification in MIDI files considering the imbalanced context

Abstract

In this paper, the problem of identifying the melodic track of a MIDI file in imbalanced scenarios is addressed. A polyphonic MIDI file is a digital score that consists of a set of tracks where usually only one of them contains the melody and the remaining tracks hold the accompaniment. This leads to a two-class imbalance problem that, unlike in previous work, is managed by over-sampling the melody class (the minority one) or by under-sampling the accompaniment class (the majority one) until both classes are the same size. Experimental results over three different music genres prove that learning from balanced training sets clearly provides better results than the standard classification process