Deep Learning on Abnormal and Evolving Graphs
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Graphs are widely used to represent complex relationships. Graph Neural Networks (GNNs) have become a powerful tool for graph learning. However, most GNNs are designed for clean and static graphs, whereas real-world graphs are often irregular—either abnormal or evolving. This thesis investigates these two underexplored yet practical scenarios.
For abnormal graphs, the focus is on graph anomaly detection, where existing methods primarily target nodes or edges within a single graph, leaving graph-level anomalies largely unstudied. Moreover, current approaches follow a one-model-one-dataset paradigm, limiting their generality. For evolving graphs, memory replay methods are a common solution, yet they fail to capture holistic information and raise privacy concerns. Additionally, the absence of task identifiers significantly degrades performance in class-incremental learning compared to task-incremental learning.
To address these issues, this thesis proposes four models. A hierarchical memory network enables graph-level anomaly detection, and a zero-shot generalist framework allows one model to generalize across datasets. For evolving graphs, a debiased lossless replay method captures more complete information, and a replay-free approach with task profiling and prompting narrows the gap between task- and class-incremental learning. Extensive experiments show that these models achieve state-of-the-art performance and provide new insights into learning on irregular graphs.
Please use this identifier to cite or link to this item: