Automated graph anomaly detection with large language models

Graph neural networks (GNNs) have emerged as powerful tools for graph anomaly detection (GAD). However, designing effective GNN architectures for GAD often demands considerable domain expertise and laborious manual tuning. Although graph neural architecture search (GNAS) has made significant progress in automating the discovery of effective deep architectures, existing GNAS methods are challenging to directly apply to GAD tasks due to the lack of a dedicated search space tailored for GAD and the difficulty in effectively incorporating domain expert knowledge into the model architecture generation process. To address these challenges, this paper proposes an automated graph anomaly detection (AutoGAD for short) framework. AutoGAD automates the generation of optimal neural network architectures through a predefined search space and an efficient search strategy. Specifically, we first design a novel search space tailored for GAD tasks based on the characteristics of the graph autoencoder framework. Then, we leverage a large language model (LLM) as the controller of GNAS, guiding the LLM to rapidly generate architectures suitable for GAD within the search space through well-designed prompts. Extensive experimental results demonstrate that AutoGAD can generate new architectures that outperform existing GAD models, and its effectiveness is consistently observed across different datasets.