Graph Anomaly Detection

• Graph anomaly detection is the process of identifying irregular graph elements, including nodes, edges, or subgraphs, that deviate from expected structural and attribute patterns.
• It employs diverse methodologies such as probabilistic models, matrix/tensor decompositions, and GNN-based as well as contrastive techniques to capture anomalies in both static and dynamic settings.
• The approach has wide applications in cybersecurity, finance, and social networks while addressing challenges like interpretability, contamination robustness, and the need for realistic evaluations.

Searching arXiv for recent and foundational papers on graph anomaly detection to ground the article in published work. Graph anomaly detection is the problem of finding graph objects whose structure, attributes, or temporal evolution do not conform to expected normal patterns. In the contemporary literature, the anomalous object may be a node, an edge, a subgraph, a whole graph in a database, or a temporally localized event, and the methodological emphasis is on modeling relational dependence rather than treating observations as independent feature vectors. GNN-based surveys further stress that anomaly detection on graphs differs from ordinary outlier detection because a node can be anomalous by attributes, by structure, or by inconsistency between the two, while graph-based time-series work adds the joint modeling of intra-variable dependency and inter-variable dependency (Kim et al., 2022, Lamichhane et al., 2024, Ho et al., 2023).

1. Conceptual scope and formal problem settings

A recurring definition across the literature is that graph anomalies are patterns in a graph that do not conform to normal patterns expected of the attributes and/or structures of the graph. This broad definition is instantiated in several distinct statistical and learning settings. In the GNN survey, anomaly detection is organized around node, edge, subgraph, and whole-graph anomalies, with static and dynamic graphs treated separately (Kim et al., 2022). The broader survey on graph-structured data uses a parallel taxonomy based on what is anomalous, what kind of graph is observed, what technique is used, and what anomaly type is sought, including outliers, dense/group anomalies, sudden anomalies, and gradual anomalies (Lamichhane et al., 2024).

The formalization varies with the task. In graph-level anomaly detection over a graph database, the input is an unlabeled graph database $G=\{G_i=(V_i,E_i)\}_{i=1}^N$, and the objective is to identify unusual graphs that differ significantly from the majority (Zhao et al., 2022). In time-series settings, a graph representation for the $j$-th observation is written as $\mathcal{G}_j=\{\mathbf{M}_j,\mathbf{A}_j\}$, while the overall graph set is $$\mathbb{G}=\{\mathcal{G}_j,\text{Sim}\{\mathcal{G}_j,\mathcal{G}_{j'}\}\}_{j\neq j'}$$; this formulation is intended to model both structure within an observation and relations between observations (Ho et al., 2023). In group-level graph anomaly detection, the task is written as $$F(\mathcal{G})\rightarrow\{\mathcal{C},\mathcal{S}\}$$, where $\mathcal{C}$ is a set of anomalous groups and $\mathcal{S}$ their scores (Ai et al., 2023).

A stricter statistical formulation appears in "Near-optimal Anomaly Detection in Graphs using Lovasz Extended Scan Statistic" (Sharpnack et al., 2013). There, anomaly detection is posed under the Gaussian normal means model as the detection of a well-connected activated region: not an arbitrary sparse set of vertices, but an elevated cluster whose boundary is small relative to the graph. This formulation makes explicit a computational-statistical tradeoff that remains central to the subject: the statistically ideal detector is often combinatorial, whereas practical detectors rely on convex relaxation, learned embeddings, or approximate search.

2. Taxonomic axes

The current literature can be summarized along a small number of recurring axes.

Axis	Categories stated in the literature	Representative sources
Anomalous object	node, edge, subgraph, whole graph, group	(Kim et al., 2022, Ai et al., 2023)
Graph regime	static, dynamic, streaming graph, streaming edge, graph time series	(Kim et al., 2022, Lamichhane et al., 2024, Ho et al., 2023)
Technique family	probabilistic/statistical, matrix/tensor decomposition, distance/similarity, GNNs	(Lamichhane et al., 2024)

This taxonomy matters because method design is tightly coupled to the anomaly object. Node-level methods dominate the recent GNN literature, especially on static attributed graphs, whereas graph-level anomaly detection in graph databases, group-level anomaly detection, and dynamic subgraph anomaly detection are much less developed (Kim et al., 2022). The graph-structured-data survey makes a similar point from a different angle: static methods remain important, but real systems increasingly require streaming graph and streaming edge formulations, where one-pass constraints, changing topology, and temporal dependence are unavoidable (Lamichhane et al., 2024).

The taxonomy also separates anomaly semantics. A node can be abnormal because its features are unusual, because its connectivity pattern is unusual, or because features and structure disagree. GAD-NR makes this distinction explicit through contextual anomalies, structural anomalies, and joint-type anomalies, and argues that many existing GAE models do well on one or two of these types but not all three (Roy et al., 2023). Group-level work extends the point further: a group may be anomalous even when the individual nodes are not node-level anomalies, because the abnormality is expressed at the level of coordinated topology patterns (Ai et al., 2023).

3. Methodological families

The survey literature identifies four broad methodological families: probabilistic/statistical methods, matrix/tensor decomposition methods, distance/similarity methods, and GNN-based methods (Lamichhane et al., 2024). The probabilistic strand remains foundational. "Near-optimal Anomaly Detection in Graphs using Lovasz Extended Scan Statistic" (Sharpnack et al., 2013) develops the generalized likelihood ratio test for well-connected activated regions and replaces the intractable graph scan with the Lovász Extended Scan Statistic (LESS), using submodularity and a poly-time graph-cut-based optimization. "Multi-Level Anomaly Detection on Time-Varying Graph Data" (Bridges et al., 2014) introduces GBTER, a community-aware generalization of BTER, and uses hierarchical probabilities and statistics to detect anomalies simultaneously at graph, subgraph, and node levels. "GraphPrints: Towards a Graph Analytic Method for Network Anomaly Detection" (Harshaw et al., 2016) represents time slices of network flow as graphs, counts graphlets, performs outlier detection on the sequence of graphlet counts, and reports false positive rates bounded by 2.84% at the time-interval level and 0.05% at the IP-level with 100% true positive rates at both.

GNN-based methods are the dominant recent family, but they differ sharply in objective. Reconstruction-based methods use graph autoencoders or variational graph autoencoders and score anomalies by reconstruction discrepancy. GAD-NR replaces direct link reconstruction by neighborhood reconstruction, reconstructing self-attributes, node degree, and neighbor-attribute distribution so that contextual, structural, and joint-type anomalies can all be captured (Roy et al., 2023). ADA-GAD argues that reconstruction-only methods suffer from Anomaly Overfitting and the Homophily Trap, and therefore pretrains autoencoders on anomaly-denoised augmentations before retraining decoders on the original graph (He et al., 2023). GEL keeps the reconstruction framework but replaces point estimates with evidential distributions, modeling node features with a Normal-Inverse-Gamma distribution and topology with Beta distributions, and adds graph uncertainty and reconstruction uncertainty to the anomaly score (Wei et al., 31 May 2025). HimNet takes a related but graph-level route by reconstructing graphs through node-level and graph-level memory modules, thereby separating local and global abnormality in a hierarchical memory-augmented graph autoencoder (Niu et al., 2023).

A second major line uses contrastive, one-class, or augmentation-based objectives. "Graph Contrastive Learning for Anomaly Detection" (Chen et al., 2021) proposes GraphCAD, which contrasts abnormal nodes with normal ones in terms of their distances to the global context and extends to a self-supervised setting through graph corruption and synthetic node labels. GLAM targets graph-level anomalies in graph databases by combining GIN encoders, Mean-pooling and MMD-pooling, a deep SVDD objective, and an unsupervised HITS-style model selection strategy (Zhao et al., 2022). DAGAD addresses labeled anomaly scarcity and class imbalance by performing representation-space data augmentation and class-wise imbalance-tailored learning with complementary classifiers (Liu et al., 2022). "GRAM: An Interpretable Approach for Graph Anomaly Detection using Gradient Attention Maps" (Yang et al., 2023) converts gradients of a trained GNN into node-level and graph-level anomaly scores, so that the anomaly score itself serves as an explanation. In the few-shot regime, MA-GAD combines graph compression, meta-learning, and deviation-based anomaly scoring for graph anomaly and subgraph anomaly detection when anomalous labels are scarce (Li et al., 9 Oct 2025).

4. Dynamic, temporal, and graph time-series anomaly detection

Dynamic graph anomaly detection is not merely the static problem repeated over time. The graph time-series survey distinguishes node, edge, subgraph, graph, and $\text{Sim}\{\cdot,\cdot\}$-level anomalies, and notes that existing methods usually address only a subset of these types (Ho et al., 2023). The broader GNN survey similarly emphasizes that dynamic graph methods often combine a graph encoder with a temporal model such as a GRU or RNN (Kim et al., 2022).

"Anomaly Detection on Graph Time Series" (Hsu, 2017) exemplifies one such design. It uses a variational recurrent neural network for anomaly detection on graph time series, models temporal correlation through the combination of recurrent neural network and variational inference, captures spatial information with graph convolutional network, incorporates external factors with a feature extractor that augments the transition of latent variables, and uses accumulative ELBO as the target function, which the abstract states is easy to extend to an on-line method. In a different formulation, "Anomaly detection in the dynamics of web and social networks" (Miz et al., 2019) defines an anomaly as a localized increase in temporal activity in a cluster of nodes. Its two-stage pipeline first filters potential anomalies from node time series and then learns a Hopfield network that reinforces links between nodes with similar temporal activity. The method is unsupervised, scalable, implemented in Apache Spark GraphX, and the paper states complexity $\mathcal{O}(NT)$ for stage 1 and $\mathcal{O}(ET)$ for stage 2.

Temporal graph anomaly detection also includes explicitly hierarchical and contextual models. "Multi-Level Anomaly Detection on Time-Varying Graph Data" (Bridges et al., 2014) is designed for streaming labeled graph sequences with persistent node identities. It models graph snapshots with GBTER, estimates node and community probabilities, and supports graph-, subgraph-, and node-level anomaly localization together with an interactive visualization workflow. This suggests a persistent theme in temporal work: detection is often inseparable from contextualization, because the relevant question is not only whether a new snapshot is anomalous, but which substructures account for the deviation.

A related temporal extension appears outside classical network settings. GETAD models map-matched trajectories on a directed road network $j$0, learns road-aware embeddings with a GAT augmented by transition probabilities, adds graph-based positional encoding, uses a Transformer decoder for sequential movement modeling, and scores anomalies with Confidence Weighted Negative Log Likelihood. On the synthetic simulation dataset it reports F1 $j$1 and PR-AUC $j$2 for constrained anomalies, while on Porto it reports PR-AUC $j$3 for constrained anomalies and $j$4 for unconstrained anomalies (Mbuya et al., 22 Sep 2025).

5. Interpretability, robustness, and evaluation

Interpretability is a persistent concern because anomaly scores are often produced by black-box representation learners. GRAM addresses this directly by defining gradient attention coefficients from $j$5 and turning them into node-level scores $j$6, so that the same object used for detection also localizes the anomalous part of the graph (Yang et al., 2023). Multi-level temporal work pursues interpretability through decomposition rather than gradients: the GBTER-based framework of (Bridges et al., 2014) supports anomaly localization at node, subgraph, and graph levels, while the web-and-social-network method of (Miz et al., 2019) treats detected communities and their temporal traces as the explanation of the event itself. Even in collider anomaly detection, interpretability is framed as control over the information available to the model through sparse graph constructions such as Laman graphs and globally rigid unique graphs (Araz et al., 24 Jun 2025).

Robustness has become equally central. GEL argues that relying solely on reconstruction error increases sensitivity to noise and overfitting, and responds by combining feature and topology reconstruction with explicit uncertainty modeling (Wei et al., 31 May 2025). ADA-GAD starts from the empirical observation that models trained on graphs with fewer anomalies exhibit higher detection performance, and therefore uses anomaly-denoised pretraining and node anomaly distribution regularization to reduce contamination effects (He et al., 2023). HimNet evaluates robustness to anomaly contamination in the training set and reports that it is consistently the best across all tested contamination levels on AIDS, BZR, DHFR, and IMDB (Niu et al., 2023). DAGAD treats labeled anomaly scarcity and class imbalance as first-order problems rather than afterthoughts, using representation-space augmentation and class-wise CE/GCE objectives (Liu et al., 2022). MA-GAD extends the robustness question to the few-shot regime by combining graph condensation with MAML-style adaptation (Li et al., 9 Oct 2025).

Evaluation practice is correspondingly heterogeneous. Node-level attributed-graph work often reports AUC and Recall@K, graph-level anomaly detection reports ROC-AUC and AP, group-level work adds Completeness Ratio (CR), trajectory anomaly detection emphasizes F1 and PR-AUC, and collider anomaly detection reports ROC AUC together with the maximum Significance Improvement Characteristic (Wei et al., 31 May 2025, Yang et al., 2023, Ai et al., 2023, Mbuya et al., 22 Sep 2025, Araz et al., 24 Jun 2025). GLAM further highlights an often neglected issue: in unsupervised graph-level anomaly detection, model selection itself is difficult because there are no labels for validation, and the paper treats unsupervised model selection as part of the anomaly-detection problem rather than a separate engineering detail (Zhao et al., 2022).

6. Applications, misconceptions, and open problems

The application range is broad. The surveys cite fraud detection, cybersecurity, finance security, e-commerce, social network analysis, computer networks, IoT, road networks, cyber-physical systems, citation and collaboration networks, healthcare monitoring, and biochemistry (Lamichhane et al., 2024, Ho et al., 2023). Specific papers add network surveillance, disease outbreak detection, activity monitoring in social networks, traffic flow data, web and social network event detection, molecular graph screening, and anomaly detection at the LHC (Sharpnack et al., 2013, Hsu, 2017, Araz et al., 24 Jun 2025). This diversity partly explains why no single anomaly definition dominates the field.

Several misconceptions recur. One is that graph anomaly detection is mainly node-level. The surveys and newer task formulations contradict this: the anomalous object may be an edge, a subgraph, a whole graph, or a group, and temporal work adds anomalous relations between graphs over time (Kim et al., 2022, Lamichhane et al., 2024, Ho et al., 2023). A second misconception is that anomalies are isolated point outliers. Statistical scan methods, Hopfield-network approaches, and group-level formulations all instead model connected regions, clusters of temporal activity, or anomalous groups whose members may individually appear normal (Sharpnack et al., 2013, Miz et al., 2019, Ai et al., 2023). A third misconception is that reconstruction error is sufficient. ADA-GAD and GEL explicitly dispute this by pointing to Anomaly Overfitting, the Homophily Trap, and noise sensitivity, while GRAM and contrastive methods shift the anomaly signal toward attribution or discrimination rather than pure reconstruction (He et al., 2023, Wei et al., 31 May 2025, Yang et al., 2023). A fourth misconception is that more connectivity or more information always improves performance. In collider anomaly detection, fully connected graphs do not perform best; the reported gains depend on specific sparse unique graph constructions and an intermediate level of subjet clustering (Araz et al., 24 Jun 2025).

Open problems are stable across surveys. The GNN survey emphasizes explainable GNNs, better anomaly-identification objectives and embedding spaces, class imbalance, heterogeneous graph anomaly detection, few-shot graph anomaly detection, and the underexplored status of edge, subgraph, graph-level, and dynamic anomalies (Kim et al., 2022). The broader survey adds adversarial robustness, graph complexity, and gradual anomalies as persistent gaps (Lamichhane et al., 2024). The graph time-series survey stresses graph construction and learning, the difficulty of jointly modeling intra-variable and inter-variable dependencies, graph augmentation for self-supervision, the lack of methods that target all anomaly types simultaneously, and problems in benchmark realism and threshold selection (Ho et al., 2023). A plausible implication is that the field is converging on a layered view of anomaly detection: anomaly scoring, uncertainty quantification, explanation, contamination robustness, and task-specific structure learning are increasingly treated as inseparable components rather than independent modules.