City University of Hong Kong
DSpace
 

CityU Institutional Repository >
3_CityU Electronic Theses and Dissertations >
ETD - Dept. of Electronic Engineering  >
EE - Doctor of Philosophy  >

Please use this identifier to cite or link to this item: http://hdl.handle.net/2031/4976

Title: Extension of self-organizing maps to structured data domain
Other Titles: Extension of self-organizing maps to structured data domain
Zi zu zhi wang luo zai jie gou xing shu ju shang de kuo zhan
自組織網絡在結構型數據上的擴展
Authors: Rahman, Md. Kaled Masukur
Department: Dept. of Electronic Engineering
Degree: Doctor of Philosophy
Issue Date: 2007
Publisher: City University of Hong Kong
Subjects: Data structures (Computer science)
Neural networks (Computer science)
Self-organizing maps
Notes: CityU Call Number: QA76.87.R33 2007
Includes bibliographical references (leaves 195-208)
Thesis (Ph.D.)--City University of Hong Kong, 2007
xiv, 208 leaves : ill. ; 30 cm.
Type: Thesis
Abstract: Self-Organizing Map (SOM) is a neural network tool to project high dimensional data into a low dimensional and topological ordered map. The application of SOM can be found in dimensionality reduction, clustering, retrieval, quantization and visualization. This thesis focuses on the developing new architectures of SOM to extend its capability from the domain of fixed-length flat-structured data to the domain of structured data. Using new SOM models, several real applications are presented that are difficult to deal with traditional SOM models. A generalized SOM model is developed named Multi-Layer Self-Organizing Map (MLSOM) to deal with general tree-structured data. MLSOM provide same functionality for tree-structured data as SOM does for flat vector type data. Traditional SOM can only deal with flat-type data that is made of fixed length input vector. The tree-structured cannot be represented by fixed length vector, and thus they cannot be dealt with traditional SOM. However, the SOM has splendid quality of dimensionality reduction and data organization. Using these two properties, MLSOM process nodes of a tree in a level-by-level fashion. In general, finding similarity between two tree data is complex and computationally expensive. On the other hand, MLSOM can process tree-structured data in a fast and efficient way by utilizing the organizing ability of SOM. Thorough experiments are presented comparing with previous SOM model in terms of unsupervised visualization, clustering and classification. A Multi-SOM model is proposed to handle applications of huge database where the data is very high dimensional and is compiled from multiple features extraction methods. Using the compression capability of SOM the Multi-SOM compresses the high dimensional data into a low dimensional “composite feature vector”, which can then be utilized for applications like clustering, visualization and retrieval. Such approach facilitates scale-free feature integration, which do not require manual scaling of feature values acquire from different source. A real world application of face identification is presented using the Multi-SOM model. Extensive experimental results are presented showing the effectiveness of the Multi-SOM approach comparing both computational demand and retrieval performance with conventional methods. A hybrid SOM (H-SOM) model is presented for classification task of tree-structured data. This model can be regarded as supervised SOM for tree-structured data in contrast to the MLSOM as unsupervised SOM. Similar to MLSOM, the H-SOM model can project very high-dimensional tree-structured data to a fixed length low dimensional space for easy classification of patterns. Basically, H-SOM consists of two SOM-based building blocks, an unsupervised block for encoding of child nodes node and a supervised layer for classification of root nodes. The H-SOM model is applied to image classification task. The classification performance using the hybrid SOM model is compared to the conventional methods. Several applications are presented using the proposed MLSOM. These include Content-Based Image Retrieval (CBIR), Document classification, and visualization and retrieval of wild flowers. The basis of all these application is the organizing capability of MLSOM for tree-structured data. Neurons on the top layer of MLSOM serve as the cluster center of tree-structured data. Thus, they provide a fast and efficient retrieval application in various domains for using tree-structured data. Detailed experimental results are presented together with the comparisons with conventional methods.
Online Catalog Link: http://lib.cityu.edu.hk/record=b2218134
Appears in Collections:EE - Doctor of Philosophy

Files in This Item:

File Description SizeFormat
fulltext.html157 BHTMLView/Open
abstract.html157 BHTMLView/Open

Items in CityU IR are protected by copyright, with all rights reserved, unless otherwise indicated.

 

Valid XHTML 1.0!
DSpace Software © 2013 CityU Library - Send feedback to Library Systems
Privacy Policy · Copyright · Disclaimer