This paper proposes FreeGAD, a novel, training-free method for Graph Anomaly Detection (GAD) to address the high deployment cost and scalability issues of deep learning-based approaches. FreeGAD uses a residual encoder with an affinity gate to generate representations that recognize outliers and uses anchor nodes as guides to compute outlier scores. Unlike existing deep learning-based GAD methods, FreeGAD demonstrates superior performance, efficiency, and scalability on various benchmark datasets without training or iterative optimization. Its development was motivated by empirical findings that the training step in existing deep learning-based GAD methods contributed less to performance than expected.
