A gradient procedure for determining clusters of relatively high point density

摘要

We propose an algorithm for determining clusters of relatively high point density. The point density is defined by a Parzen estimate of the underlying probability density. A Gaussian bump is chosen for the Parzen window function. The algorithm puts points which can be connected by gradient lines to a maximum x0 of the point density, into the same (gradient) cluster (around x0). For this task a gradient procedure with step control is employed. We compare the procedure's convergence properties and computational expenses to those of other procedures for determining gradient clusters. Notes for choosing optimal standard deviations of the Gaussian bump are given.