ARC+: An Improved Generalist Graph Anomaly Detector

Node anomaly detection in graphs presents unique challenges, particularly when deploying methods across datasets with different structures and attributes. Unlike most state-of-the-art methods, which require extensive training to each specific dataset, ARC (A Generalist Graph Anomaly Detector with In-Context Learning) introduces a novel training-free approach aimed at providing a generalized solution. By removing the dependency on dataset-specific training, ARC offers significant advantages, including reduced computational overhead, rapid adaptability to new datasets and scalability to graphs. This thesis extends the ARC model by introducing three small modifications: the first focuses on feature alignment, the second and third introduce two different classifiers to the model. These changes have been evaluated on multiple datasets, leading to observed performance improvements. However, refining the model is challenging because ARC is designed to work across different datasets without fine-tuning, making it hard to optimize for a specific case without affecting its general applicability.