That Doesn't Belong There

Big DataBig Data

Posted: May 27, 2024

That Doesn't Belong There

Anomalies in graphs signify deviations from anticipated patterns, either in attributes or structural elements. In recent years, Graph Neural Networks (GNNs) have emerged as a focal point of research, demonstrating remarkable efficacy in handling complex machine learning tasks such as node classification, link prediction, and graph classification.

This is largely attributed to their adeptness in acquiring comprehensive graph knowledge through message-passing mechanisms.

To address the critical task of graph anomaly detection, GNN-based methodologies leverage graph attributes and structures to meticulously assess and score anomalies. This survey undertakes a comprehensive examination of recent advancements in the field of detecting graph anomalies using GNN models.

Analysis categorizes these methodologies based on graph types, distinguishing between static and dynamic, anomaly types encompassing nodes, edges, subgraphs, and entire graphs, and network architectures, including graph autoencoders and graph convolutional networks.

Notably, this survey marks the inaugural endeavor in providing an extensive review of graph anomaly detection techniques employing GNNs.