[Paper Review] Graph Anomaly Detection with Graph Neural Networks: Current Status and Challenges

Hwan Kim, Byung Suk Lee+2 more|arXiv (Cornell University)|Sep 29, 2022

Advanced Graph Neural Networks4 citations

TL;DR

This survey provides a comprehensive review of graph neural network (GNN)-based methods for graph anomaly detection, categorizing approaches by graph type (static/dynamic), anomaly type (node/edge/subgraph/whole graph), and architecture (e.g., graph autoencoder, GCN). It identifies key challenges such as explainability, class imbalance, heterogeneous graphs, and few-shot learning, offering a foundational reference for future research in GNN-driven graph anomaly detection.

ABSTRACT

Graphs are used widely to model complex systems, and detecting anomalies in a graph is an important task in the analysis of complex systems. Graph anomalies are patterns in a graph that do not conform to normal patterns expected of the attributes and/or structures of the graph. In recent years, graph neural networks (GNNs) have been studied extensively and have successfully performed difficult machine learning tasks in node classification, link prediction, and graph classification thanks to the highly expressive capability via message passing in effectively learning graph representations. To solve the graph anomaly detection problem, GNN-based methods leverage information about the graph attributes (or features) and/or structures to learn to score anomalies appropriately. In this survey, we review the recent advances made in detecting graph anomalies using GNN models. Specifically, we summarize GNN-based methods according to the graph type (i.e., static and dynamic), the anomaly type (i.e., node, edge, subgraph, and whole graph), and the network architecture (e.g., graph autoencoder, graph convolutional network). To the best of our knowledge, this survey is the first comprehensive review of graph anomaly detection methods based on GNNs.

Motivation & Objective

• To provide a systematic and comprehensive review of recent advances in graph anomaly detection using graph neural networks (GNNs).
• To categorize existing GNN-based methods according to graph type (static vs. dynamic), anomaly type (node, edge, subgraph, whole graph), and network architecture.
• To identify and analyze key challenges in GNN-based graph anomaly detection, including explainability, class imbalance, and handling heterogeneous or dynamic graphs.
• To highlight underexplored research directions such as few-shot learning and anomaly detection in complex graph structures.
• To serve as a foundational reference for researchers aiming to advance GNN-based graph anomaly detection in real-world applications.

Proposed method

• Classify GNN-based graph anomaly detection methods along three axes: graph type (static/dynamic), anomaly type (node/edge/subgraph/whole graph), and network architecture (e.g., GCN, graph autoencoder).
• Review representative models such as graph autoencoders and GCN-based architectures that leverage both node features and graph topology for anomaly scoring.
• Analyze the use of reconstruction loss and distance-based functions in GNNs for identifying anomalies via embedding space deviation.
• Evaluate the role of message passing mechanisms in capturing structural and feature-based anomalies simultaneously.
• Propose the use of hyperbolic or flexible embedding spaces to better model complex graph geometry and improve anomaly detection robustness.
• Discuss strategies such as data augmentation and meta-learning to address class imbalance and few-shot anomaly detection in low-data regimes.

Experimental results

Research questions

• RQ1 How do GNN-based methods effectively detect anomalies by integrating both structural and feature-based patterns in graphs?
• RQ2 What are the key architectural differences and performance trade-offs among GNN models used for graph anomaly detection across static and dynamic graphs?
• RQ3 Why is explainability a critical challenge in GNN-based graph anomaly detection, and what are the current limitations in this area?
• RQ4 How does class imbalance between normal and anomalous instances affect GNN-based anomaly detection, and what mitigation strategies are effective?
• RQ5 What are the open challenges in detecting anomalies in heterogeneous or dynamic graphs using GNNs, and what future research directions are most promising?

Key findings

• The majority of current research focuses on detecting node anomalies in static graphs, with limited work on edge, subgraph, or whole-graph anomalies.
• GNN-based models effectively combine structural and feature information through message passing, enabling joint detection of anomalies across multiple modalities.
• Reconstruction-based loss functions in GNNs are vulnerable to noise and may fail when anomalies and normal instances are far apart in embedding space.
• There is a significant lack of research on explainable GNNs for graph anomaly detection, despite the need for interpretable anomaly explanations.
• Class imbalance remains a major challenge, with few effective data augmentation or sampling strategies proposed specifically for graph anomaly detection.
• Few-shot graph anomaly detection with GNNs is underexplored, though recent efforts show promise using meta-learning and few labeled anomalies.