Islands of Gaussians: The Self Organizing Map and Gaussian Music Similarity Features.

Multivariate Gaussians are of special interest in the MIR field of automatic music recommendation. They are used as the de facto standard representation of music timbre to compute music similarity. However, standard algorithms for clustering and visualization are usually not designed to handle Gaussian distributions and their attached metrics (e.g. the Kullback-Leibler divergence). Hence to use these features the algorithms generally handle them indirectly by first mapping them to a vector space, for example by deriving a feature vector representation from a similarity matrix. This paper uses the symmetrized Kullback-Leibler centroid of Gaussians to show how to avoid the vectorization detour for the Self Organizing Maps (SOM) data visualization algorithm. We propose an approach so that the algorithm can directly and naturally work on Gaussian music similarity features to compute maps of music collections. We show that by using our approach we can create SOMs which (1) better preserve the original similarity topology and (2) are far less complex to compute, as the often costly vectorization step is eliminated.