A computational study of several relocation methods for k-means algorithms
 
 
 

Summary
The purpose of this paper is to report and discuss the results of an empirical investigation of several techniques used by k-means algorithms (based on the Friedman-Rubin approach) to move entities from one cluster to another. Most of these procedures differ basically in the number of criterion evaluations required to reach an optimum and the accuracy of this optimum. The prime objective of the current research study has been to establish the relative merits of seventeen combinatorial passes by comparing them across a variety of artificial data sets. The experimental results suggest that a direct and efficient search which moves down the steepest permissible direction globally outperforms both simple and more sophisticated reassignment methods in terms of grouping efficacy and numerical efficiency.
 

keywords
non-hierarchical classification, iterative partitioning, combinatorial optimization
Pattern recognition, 2003, Vol. 36, n.12, 2955-2966. 

Back to main publications	Home page	Adobe pdf version