E-Thesis 182 views 32 downloads
Deep Learning on Graphs: Directed Graphs, Edge Structures and Graph Estimation / MICHAEL KENNING
Swansea University Author: MICHAEL KENNING
-
PDF | E-Thesis – open access
Copyright: The Author, Michael P. Kenning, 2023. Distributed under the terms of a Creative Commons Attribution 4.0 International License (CC BY 4.0).
Download (8.58MB)
DOI (Published version): 10.23889/SUthesis.65374
Abstract
In the last decade and a half, machine learning has been refounded on a class of techniques called deep learning. The earliest, most prominent techniques of deep learning were restricted in their application to regularly structured domains. A new set of techniques, broadly referred to as geometric d...
| Published: |
Swansea, Wales, UK
2023
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Xie, Xianghua. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa65374 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Abstract: |
In the last decade and a half, machine learning has been refounded on a class of techniques called deep learning. The earliest, most prominent techniques of deep learning were restricted in their application to regularly structured domains. A new set of techniques, broadly referred to as geometric deep learning, extends the application of deep learning approaches to irregular domains, in particular the use of the graph. A graph is an effective means of representing irregular relations between discretely sampled points; its use has its attendant research challenges that has brought about a flourishing field of research. In this work we investigate three of those challenges, namely learning on directed graphs, one of the many variants of the graph; learning on the edge-structure of graphs; and graph estimation, i.e. the estimation of graph structure from the data itself. In the first chapter of our work, we consider the challenge of learning on the edge structure of a graph in application to a datacentre and present a convolution technique for the edge-structure of a directed graph representing a datacentre. In the second chapter of our work, we present a strategy to estimate two complementary graphs, the long-term or static graph and short-term or dynamic graph from combinations of temporal, cyclical data. Additionally, we propose an attention-based convolution for directed graphs that factorises neighbourhood signals into in- and out-flows. In the third and final chapter of our work, we draw on the work of the previous two chapters and design a graph estimation strategy to learn the complementary structures of molecular graphs. For this purpose, we define a new kind of graph for the estimation, which we call the graph complement, and use it in predicting molecular properties by incorporating intramolecular forces not present in the original graph. The structure thus learned is used to propagate vertex and edge signals on a directed graph. The work is concluded with a reflection on the contributions of the thesis and prospective areas of research in the field of graph deep learning. |
|---|---|
| Keywords: |
Deep learning, graph deep learning, directed graph, edge learning, graph estimation |
| College: |
Faculty of Science and Engineering |
| Funders: |
Swansea University Research Excellence Scholarship |