Date of Award

2010

Thesis Type

PhD

Document Type

Thesis

Divisions

Faculty of Science

Department

Institute of Mathematical Sciences

Institution

Universiti Malaya

Abstract

This thesis develops mixture model clustering algorithms scalable to data sets that do not fit into the computer memory buffer. Two algorithms, FlexClust and FlexClustMix, are proposed for clustering very large continuous and mixed data sets respectively. The basic framework of the algorithms is to scale down data incrementally using data compression, and incorporates the later compressed data into the current model with the ability to recover new clusters that have been missed out in the earlier compressed data. It consists of three modules: 1) incremental compression, 2) detecting change in cluster structure, and 3) update of current model. In FlexClust, Gaussian mixture model is used to compress data. The incorporation of the incrementally compressed data into the current model is done through the proposed bi-Gaussian mixture model. In FlexClustMix, a mixture model for mixed data, the Gaussian location model, which speeds up parameters estimation, is proposed for compression. The incorporation of the incrementally compressed data into the current model is done through the proposed bi-Gaussian location model. A model selection criterion, modified Bayes factor (MBF), is proposed to detect changes in clusters structure due to the incrementally added data and to recover any small clusters that have been missed out in the initial sample. FlexClust and FlexClustMix are tested over very large continuous and mixed data sets respectively and the results are promising.

Initial

khm

Additional Information

Thesis (PhD) - Faculty of Science, Universiti Malaya, 2010.

Download

Share

COinS