Anomaly Detection for Large-Scale Text-Attributed Graphs and Its Applications

Anomaly detection in graph data has seen widespread applications across numerous domains, such as fraud detection in financial networks, bot account identification in social media, and malicious review detection on e-commerce websites. The majority of existing anomaly detection approaches for graph data are based on graph neural networks, whose performance heavily relies on a large number of high-quality labeled samples. However, in real-world scenarios, graph data are typically large in scale, rendering anomaly labels hard to obtain and the annotation process labor-intensive and highly costly. Additionally, real graphs often present as Test-Attributed Graphs (TAGs), wherein nodes/edges are endowed with rich textual content embodying crucial information for anomaly detection. Unfortunately, to our knowledge, none of existing methods have effectively leveraged both the textual and structural information therein for enhanced anomaly detection performance.

To address the challenges of dual data modalities, label scarcity, and the massive scale of TAGs in current graph anomaly detection tasks, this proposal seeks to design and implement an adaptive anomaly detection framework tailored for large-scale, complex, and dynamic TAGs. This project intends to achieve this goal through four key research tasks: (i) integrating outlier detection algorithms with large language models to enable efficient training sample annotation and synthesis; (ii) designing feature encoding and representation learning models for dual-modal data comprising text and graph structures; (iii) implementing rapid and accurate detection result correction and model calibration; and (iv) developing a series of efficiency optimization techniques for large-scale TAGs to facilitate model training and inference.