Data driven point packing for fast clustering

Modern data acquisition has forced the field of large data on the scientific community. This papers gives a rapid technique for clustering data. The technique is based on an off-line process for packing points chosen from a data space. Once the off-line process has been run, the clustering may be re-run on different data sets of the same type in linear time. The clustering takes the form of a Voronoi tiling of the data space with the tile centres being the elements of the point packing. The data items within each tile form the clusters. The evolutionary algorithm is an adaptation of one, based on the Conway crossover operator, that has been used to create error correcting codes over the Levenstein metric; the tile centres are a form of code, but over the Euclidean metric. The technique generalizes smoothly to other metric spaces and may be used on any type of data for which a distance metric can be devised. The data set used in this study captures information about codon usage bias in human genes. The clustering is validated by looking for GO term over- representation in the clusters, with significant results.