Open Collections

Liu, Xiaoou

University of British Columbia

This thesis studies the effectiveness of Graph Neural Networks (GNNs) for node classification. We conduct systematic experiments on several representative deep-learning models for graph data, using training data generated from the Stochastic Block Model (SBM) and the theoretical results on the fundamental limits of this model as guidance in the design of our experiments. While GNNs are widely believed to be powerful learning models for graph data, our empirical findings suggest that they do not necessarily outperform other machine-learning-based methods and traditional algorithms for node classification. In particular, we observe that GNN-based methods fail to exploit the information from labeled nodes in semi-supervised learning settings. We propose an effective data augmentation method to enhance GNN-based methods by making better use of labeled information in the training data. Our experiments using synthetic data from SBMs and real-world datasets demonstrate that our method can significantly enhance the capabilities of GNN models and notably improve their performance for node classification. Additionally, in the context of unsupervised learning, we discuss the possibility of incorporating other types of side information into GNNs that may exist in multiplex network data.

Text

2024-06-17

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/88475

Computer Science

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/