Local Optima in K-Means Clustering

"The study of the properties of local optimality in K-means clustering is pursued. In doing so, it is shown that several of the commercial software packages prove to be inadequate in their treatment of the K-means algorithm, resulting in the proposal of an alternative method based on several thousand initializations, which is imbedded in a MATLAB m-file. The further developments of this dissertation are four-fold: (a) a comprehensive cluster generation method based on distributional theory and probability is developed; (b) the properties of local optimality are related to a cluster recovery criterion to develop a test that is able to distinguish between ""good"" and ""bad"" cluster solutions; (c) a method of consensus analysis for K-means clustering is proposed and extended to within-cluster standardization; and (d) a lower bound for the K -means criterion function is derived, and based on the lower bound, another (more powerful) test is developed to determine the quality of a given cluster solution."

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Psychology

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Steinley, Douglas Lee

Contributors dc:contributor

• Hubert, Lawrence J.

• Psychology, Psychometrics

Language dc:language

eng

Identifier

(MiAaPQ)AAI3131029

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/82057