Year

2015

Degree Name

Doctor of Philosophy

Department

School of Computing and Information Technology

Abstract

While the application of machine learning algorithms to practical problems has been expanded from fixed sized input data to sequences, trees or graphs input data, the composition of learning system has developed from a single model to integrated ones. Recent advances in graph based learning algorithms include: the SOMSD (Self Organizing Map for Structured Data), PMGraphSOM (Probability Measure Graph Self Organizing Map,GNN (Graph Neural Network) and GLSVM (Graph Laplacian Support Vector Machine). A main motivation of this thesis is to investigate if such algorithms, whether by themselves individually or modified, or in various combinations, would provide better performance over the more traditional artificial neural networks or kernel machine methods on some practical challenging problems. More succinctly, this thesis seeks to answer the main research question: when or under what conditions/contexts could graph based models be adjusted and tailored to be most efficacious in terms of predictive or classification performance on some challenging practical problems? There emerges a range of sub-questions including: how do we craft an effective neural learning system which can be an integration of several graph and non-graph based models? Integration of various graph based and non graph based kernel machine algorithms; enhancing the capability of the integrated model in working with challenging problems; tackling the problem of long term dependency issues which aggravate the performance of layer-wise graph based neural systems. This thesis will answer these questions.

Recent research on multiple staged learning models has demonstrated the efficacy of multiple layers of alternating unsupervised and supervised learning approaches. This underlies the very successful front-end feature extraction techniques in deep neural networks. However much exploration is still possible with the investigation of the number of layers required, and the types of unsupervised or supervised learning models which should be used. Such issues have not been considered so far, when the underlying input data structure is in the form of a graph. We will explore empirically the capabilities of models of increasing complexities, the combination of the unsupervised learning algorithms, SOM, or PMGraphSOM, with or without a cascade connection with a multilayer perceptron, and with or without being followed by multiple layers of GNN. Such studies explore the effects of including or ignoring context. A parallel study involving kernel machines with or without graph inputs has also been conducted empirically.

FoR codes (2008)

0801 ARTIFICIAL INTELLIGENCE AND IMAGE PROCESSING, 0803 COMPUTER SOFTWARE, 0899 OTHER INFORMATION AND COMPUTING SCIENCES

Download

Share

COinS
 

Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.