Hierarchical density-based clustering in high-dimensional spaces using topographic maps

A novel way to perform hierarchical, divisive clustering is outlined in this paper. Rather than exhaustively subdividing the complete data set, a density estimate, obtained using topographic maps, is analyzed at every level in the hierarchy in order to determine the number of clusters and to divide the data into new subsets to be analyzed at the next level. Our algorithm is illustrated using a real-world example comprising high-dimensional music data (spectrograms). The different levels of similarity one intuitively perceives in the music signal, correspond to the clustering results found by the algorithm.