Constructing Proximity Graphs to Explore Similarities in Large-Scale Melodic Datasets

This paper investigates the construction of proximity graphs in order to allow users to explore similarities in melodic datasets. A key part of this investigation is the use of a multilevel framework for measuring similarity in symbolic musical representations. The basis of the framework is straightforward: initially each tune is normalised and then recursively coarsened, typically by removing weaker off-beats, until the tune is reduced to a skeleton representation with just one note per bar. Melodic matching can then take place at every level: the multilevel matching implemented here uses recursive variants of local alignment algorithms, but in principle a variety of similarity measures could be used. The multilevel framework is also exploited with the use of early termination heuristics at coarser levels, both to reduce computational complexity and, potentially, to enhance the matching qualitatively. The results of the matching algorithm are then used to construct proximity graphs which are displayed as part of an online interface for users to explore melodic similarities within a corpus of tunes.