A vector field approach to cluster analysis

摘要

An efficient computer method which uses an extension of the familiar gravitational field to find clusters of multidimensional data is suggested. The manipulation of a single parameter, r, permits one's perspective of the data to range from the locally sensitive (where each datum is a cluster) to the globally sensitive (where the entire sample set is regarded as one cluster). The number of clusters and their locations are determined by a choice of r. The program finds clusters by converging on the nodes of the field in decreasing steps. Once located, the field is modified so that the known nodes are effectively precluded from further consideration.