Generative Models Driven Graph Outlier Detection

Graph data are pervasive across various domains, including social networks, biological networks, and communication systems. The detection of outliers in graph data—substructures that significantly deviate from the norm—is crucial for uncovering fraudulent activities, network vulnerabilities, and novel patterns. However, traditional outlier detection techniques often fall short of effectively modeling the complex non-Euclidean graph data. Recently, generative models have exhibited extraordinary performance on image and language data, but their capabilities in graph outlier detection remain largely underexplored. In this dissertation, I systematically investigate the capabilities of generative models in the context of graph outlier detection, including four published works and one ongoing work. Specifically, first two works introduce a comprehensive graph outlier detection library and a benchmark on existing graph outlier detection algorithms. The third work presents a diffusion model–based data augmentation for addressing class imbalance in graph outlier detection. The fourth work explores scalable global spatiotemporal attention in graph Transformers for graph outlier detection. The last work focuses on the fake news detection capabilities of large language models and constructs a large-scale real-world text-attributed graph dataset for graph outlier detection.

Author dc:creator

• Kay Zekuan Liu (23291320)

• Computer Science

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451065.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451065