Empirically defined regions of influence for clustering analyses

摘要

We present a simple but effective clustering algorithm for non-hierarchical k-dimensional data which is also useful for pattern recognition applications. The method uses a visually-empirical region of influence (VERI) that we have discovered. Our approach enables us to use human cluster judgments as the cluster criteria, and to extend these criteria in a natural way to k-dimensional data. The VERI algorithm requires no user input (e.g. the number of clusters in the final result) or user adjustments beyond providing the data itself. The algorithm computes clusterings based on the local k-dimensional neighbors of each point, and thus handles arbitrary numbers of clusters and arbitrary global cluster shapes. We demonstrate that the method works well for a variety of 2D cluster configurations which popular methods cannot treat (despite requiring the number of clusters as input), and illustrate that the performance is maintained for k-dimensional problems. The application of the method to pattern recognition problems is outlined, and efficient implementations of the VERI algorithm are presented.