Graph Representation Learning for Graph-level Classification and Anomaly Detection
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Graph, due to its specific relation reservation ability, has become one of the most popular data storage modes in numerous. Graph representation learning is a crucial topic in graph data mining. Increasing attention is attracted by the graph neural networks (GNNs) as a result of their remarkable success in learning informative graph representations. Improving the expressiveness of GNN can help learn information-richer graph representations and thus is always a significant research goal. Besides, since the data collection is a costly process, how to learn expressive graph representations with limited supervision information is also important. This thesis leverages three popular and practical graph-level classification tasks to formulate the circumstances with different amounts of supervision information and proposes three models to learn more expressive graph representations to solve them. In detail, our thesis makes the following contributions: (i) We introduce a collective node and graph-level structural information harnessed model to improve the expressiveness of GNN. The proposed method significantly outperforms competing methods in graph classification task and is more generalized to out-of-distribution graphs, enabling it to be applied to real-world application in industry for higher accuracy and coping with unknown certainty in context. Besides, the proposed method is resource- and time-efficient, enabling the method to be applied to more platforms in industry; (ii) We propose a novel deep multi-scale oversampling framework and its instantiation to address the imbalanced graph classification problem. This is the first work that takes account of both within and between graph information to learn graph representations for imbalanced graph classification. This research improves the performance for data-insufficiency tasks and can save large amount of human resource in data collection; (iii) We formulate the graph-level anomaly detection problem as the task of detecting locally- or globally-anomalous graphs and empirically verify the presence of these two types of graph anomalies in real-world datasets. Then we introduce the first approach and its instantiation specifically designed to effectively detect both types of anomalous graphs. The proposed method performs significantly better and can be trained much more sample-efficiently and with more robustness when compared with its advanced counterparts. This research would be of great importance to varying applications in industry, e.g., omitting numerous experiments to identify toxic molecules from a set of chemical compounds, reducing human trials in recognizing drugs with severe side-effects or preventing the loss of millions of dollars by detecting fraud communities.
Please use this identifier to cite or link to this item: