Robustness, Heterogeneity and
Structure Capturing for Graph Representation Learning and its Application

Abstract

Graph neural networks (GNNs) are potent methods for graph representation learn- ing (GRL), which extract knowledge from complicated (graph) structured data in various real-world scenarios. However, GRL still faces many challenges. Firstly GNN-based node classification may deteriorate substantially by overlooking the pos- sibility of noisy data in graph structures, as models wrongly process the relation among nodes in the input graphs as the ground truth. Secondly, nodes and edges have different types in the real-world and it is essential to capture this heterogeneity in graph representation learning. Next, relations among nodes are not restricted to pairwise relations and it is necessary to capture the complex relations accordingly. Finally, the absence of structural encodings, such as positional information, deterio- rates the performance of GNNs. This thesis proposes novel methods to address the aforementioned problems:
1. Bayesian Graph Attention Network (BGAT): Developed for situations with scarce data, this method addresses the influence of spurious edges. Incor- porating Bayesian principles into the graph attention mechanism enhances robustness, leading to competitive performance against benchmarks (Chapter 3).
2. Neighbour Contrastive Heterogeneous Graph Attention Network (NC-HGAT): By enhancing a cutting-edge self-supervised heterogeneous graph neural net- work model (HGAT) with neighbour contrastive learning, this method ad- dresses heterogeneity and uncertainty simultaneously. Extra attention to edge relations in heterogeneous graphs also aids in subsequent classification tasks (Chapter 4).
3. A novel ensemble learning framework is introduced for predicting stock price movements. It adeptly captures both group-level and pairwise relations, lead- ing to notable advancements over the existing state-of-the-art. The integration of hypergraph and graph models, coupled with the utilisation of auxiliary data via GNNs before recurrent neural network (RNN), provides a deeper under- standing of long-term dependencies between similar entities in multivariate time series analysis (Chapter 5).
4. A novel framework for graph structure learning is introduced, segmenting graphs into distinct patches. By harnessing the capabilities of transformers and integrating other position encoding techniques, this approach robustly capture intricate structural information within a graph. This results in a more comprehensive understanding of its underlying patterns (Chapter 6).