With a long history of traditional Graph Anomaly Detection (GAD) algorithms and recently popular Graph Neural Networks (GNNs), it is still not clear (1) how they perform under a standard comprehensive setting, (2) whether GNNs outperform traditional algorithms such as tree ensembles, and (3) their efficiency on large-scale graphs. In response, we present GADBench -- a comprehensive benchmark for supervised anomalous node detection on static graphs. GADBench provides a thorough comparison across 23 distinct models on ten real-world GAD datasets ranging from thousands to millions of nodes ($\sim$6M). Our main finding is that tree ensembles with simple neighborhood aggregation outperform all other baselines, including the latest GNNs tailored for the GAD task. By making GADBench available as an open-source tool, we offer pivotal insights into the current advancements of GAD and establish a solid foundation for future research. Our code is available at https://github.com/squareRoot3/GADBench.

翻译：具有传统的图异常检测（GAD）算法和近期流行的图神经网络（GNNs）的悠久历史，目前仍不清楚：（1）它们在标准综合设置下的性能如何；（2）GNNs是否优于传统算法（如树集成方法）；（3）它们在大规模图上的效率。为此，我们提出了GADBench——一个针对静态图上监督异常节点检测的综合性基准。GADBench在十个真实世界GAD数据集上对23种不同模型进行了全面比较，这些数据集规模从数千节点到数百万节点（约600万）。我们的主要发现是：采用简单邻域聚合的树集成方法优于所有其他基线模型，包括专门为GAD任务定制的最新GNNs。通过将GADBench作为开源工具发布，我们为当前GAD的进展提供了关键见解，并为未来研究奠定了坚实基础。我们的代码可在https://github.com/squareRoot3/GADBench获取。