Graph-Level Anomaly Detection (GLAD)

• Graph-level anomaly detection is the process of scoring entire graphs based on deviations in topology, node attributes, or global structural semantics.
• Techniques include one-class learning, reconstruction-based methods, and spectral discrimination to capture both local and global anomalies.
• Recent approaches incorporate explainability by providing explanatory subgraphs or prototype comparisons to make anomaly detection more interpretable.

Graph-level anomaly detection (GLAD) is the task of identifying entire graphs in a collection that deviate from the majority in topology, node attributes, or both. In the common formulation, each sample graph is written as $G=(A,X)$ or $\mathcal{G}=(\mathcal{V},\mathcal{E})$, and the objective is to learn an anomaly scoring function $f:\mathcal{G}\rightarrow \mathbb{R}$ so that graphs that conform less to the normal graph population receive higher scores. A central distinction in the literature is between locally anomalous graphs, whose abnormality is induced by unusual nodes or local substructures, and globally anomalous graphs, whose overall graph-level semantics or structure are deviant. More recent work extends this formulation to explainable settings, where the detector outputs both an anomaly score and an explanatory subgraph (Ma et al., 2021, Niu et al., 2023, Liu et al., 2023).

1. Problem boundaries and task variants

GLAD is distinct from node-level anomaly detection and subgraph anomaly detection. Node-level methods assign anomaly scores to individual vertices inside one or more graphs, whereas graph-level methods assign a score or label to each whole graph. Subgraph anomaly detection targets suspicious local regions; graph-level anomaly detection instead asks whether an entire graph is abnormal. This granularity distinction is explicit in few-shot GLAD work, which separates node-level, subgraph-level, and graph-level outputs and emphasizes whole-graph decisions as the defining property of GLAD (Li et al., 9 Oct 2025).

The task boundary is also important because the acronym “GLAD” is overloaded. “GLAD: Content-aware Dynamic Graphs For Log Anomaly Detection” addresses anomalous relations in dynamic log graphs and interval-level log anomalies, not the standard graph-set problem of scoring entire graphs in a dataset (Li et al., 2023). Likewise, “GLAD (Graph Learning for Anomaly Detection)” in citation networks targets anomalous citation relationships using node semantics, edge semantics, and citation purpose, rather than graph-level anomaly detection over a set of graphs (Liu et al., 2022). In the graph anomaly literature proper, GLAD usually denotes the graph-set task in which normality is learned across multiple graphs and transferred to unseen graphs.

2. Core learning paradigms

A major early line of work models normality through one-class representation learning and distillation. GLocalKD learns rich global and local normal pattern information by training one GNN to predict another randomly initialized GNN at both graph and node levels; its anomaly score is the sum of graph-level and node-level prediction errors (Ma et al., 2021). HimNet instead adopts a graph autoencoder with hierarchical memory: a node memory module models fine-grained internal graph interactions for locally abnormal graphs, and a graph memory module models holistic normal patterns for globally abnormal graphs. The anomaly score is the joint reconstruction and approximation error, and the paper reports robustness to anomaly contamination (Niu et al., 2023).

A second line combines one-class compactness with self-supervised transformation learning. OCGTL uses an ensemble of $K+1$ GNN feature extractors, an OCC term that maps normal graphs into a compact region, and a graph transformation learning term that enforces diversity and relevance among learned views. The same objective is used as the anomaly score at inference time. This design is presented as addressing two specific failures of earlier one-class graph methods: hypersphere collapse and performance flip (Qiu et al., 2022).

A third line begins from reconstruction, but recent work has challenged the standard assumption that a high mean reconstruction error is sufficient. MuSE shows that graph autoencoder methods can suffer from reconstruction flip, where an unseen graph can be reconstructed better than a training graph even when it is intuitively more anomalous. The proposed remedy is to treat reconstruction errors as features and summarize them by multiple statistics, specifically mean and standard deviation, over node-feature and adjacency reconstruction errors, yielding a 4-dimensional graph representation for a downstream one-class classifier. The paper argues that mean alone is too coarse because it ignores the shape and spread of the reconstruction-error distribution (Kim et al., 2024).

3. Spectral, geometric, and long-range extensions

A prominent recent direction reframes GLAD as a problem of spectral discrimination. RQGNN reports that normal and anomalous graphs have significantly different accumulated spectral energy, and proves that this quantity is represented by the Rayleigh quotient,

$RQ(x,L)=\frac{x^TLx}{x^Tx}.$

Its architecture couples an explicit Rayleigh Quotient Learning branch with a Chebyshev Wavelet GNN and an RQ-guided pooling operator, thereby combining global spectral signatures with multi-scale spectral representations (Dong et al., 2023).

Several subsequent models broaden this spectral perspective. HC-GLAD constructs motif-based hypergraphs to capture high-order node-group interactions and embeds both graphs and hypergraphs in hyperbolic space via the hyperboloid model, arguing that hierarchy and power-law structure are indispensable factors in unsupervised graph-level anomaly detection (Fu et al., 2024). GLADformer combines a Graph Transformer module with global spectrum enhancement and a band-pass spectral GNN message-passing module, explicitly targeting global anomaly representations and spectral characteristics in labeled graph settings (Xu et al., 2024). GLADMamba introduces selective state space models into unsupervised GLAD through a View-Fused Mamba module and a Spectrum-Guided Mamba module based on the Rayleigh quotient, with the stated goal of capturing long-range dependencies efficiently while preserving explicit spectral cues (Fu et al., 23 Mar 2025).

Taken together, these methods move the field beyond strictly spatial message passing. They treat anomalies not only as deviations in neighborhood aggregation, but also as shifts in spectral energy, high-order group structure, hierarchy, and long-range dependency patterns.

4. Scoring functions, failure modes, and evaluation

The literature increasingly treats the anomaly score function as a first-class design object rather than a by-product of representation learning. MuSE’s reconstruction-flip analysis is an explicit critique of the widespread rule “high mean reconstruction error = anomaly,” showing empirically and theoretically that this assumption can fail when test graphs share the same primary pattern as training graphs but with stronger pattern strength (Kim et al., 2024). OCGTL makes an analogous point for one-class objectives: compactness alone admits trivial solutions, and a useful anomaly score must remain discriminative under training dynamics that otherwise induce collapse or performance inversion (Qiu et al., 2022).

Two additional stressors recur in evaluation: class imbalance and training contamination. IGAD-CF treats anomalies as the minority class, constructs anomalous samples through counterfactual augmentation, complements node attributes with degree attributes, and reports the best AUC on 7 of 8 public datasets, with all reported standard deviations below 5% (Wang et al., 2024). At a broader level, UB-GOLD unifies unsupervised GLAD and graph-level OOD detection as generalized graph-level OOD detection, providing a benchmark with 35 datasets spanning four scenarios and comparing 18 representative methods. Its multi-dimensional analyses report that end-to-end methods are generally strongest, near-OOD is harder than far-OOD, and performance degrades as contamination increases in the training set (Wang et al., 2024).

This benchmark perspective changes how GLAD results are interpreted. Rather than treating anomaly detection and graph-level OOD detection as isolated subfields, it places both under the same generalized detection problem and evaluates robustness, efficiency, OOD sensitivity, and contamination tolerance in a common protocol.

5. Interpretability and explanation

Explainability has become a distinct GLAD subproblem. SIGNET defines explainable GLAD as learning a function $f:\mathbb{G}\to(\mathbb{R},\mathbb{G})$ that returns both an anomaly score and an explanatory subgraph. Its central mechanism is the multi-view subgraph information bottleneck (MSIB): bottleneck subgraphs are extracted from an input graph and its dual hypergraph, their representations are aligned by cross-view mutual information, and the anomaly score is the negative mutual information between the two view-specific bottleneck representations. The extracted bottleneck subgraphs serve as rationales. On six datasets with ground-truth rationales, the paper reports average gains of $+27.89\%$ in NX-AUC and $+8.99\%$ in EX-AUC over the best baselines (Liu et al., 2023).

ProtoGLAD advances a different explanation philosophy: contrast with normal prototypes. Instead of explaining a graph only by highlighting suspicious internal substructures, it discovers multiple normal prototype graphs and their clusters by a point-set kernel, then scores a graph by its similarity to the nearest normal cluster. An anomaly is explained through explicit contrast with its nearest normal prototype graph from the dataset, not with an abstract latent vector. The paper reports the best average rank, 1.25, on eight real-world datasets while emphasizing that concrete prototype graphs provide better human-interpretable explanations than latent prototypes or post-hoc rationales without a normal reference (Zhao et al., 11 Feb 2026).

The field therefore now contains two complementary explanation strategies: rationale extraction, which identifies substructures that drive abnormality, and prototype-based contrast, which situates abnormality relative to a concrete normal graph.

6. Applications, supervision regimes, and emerging frontiers

GLAD has been studied across biochemistry, social networks, and domain-specific graph construction pipelines. RQGNN motivates the task through applications such as cancer diagnosis and enzyme prediction, while many benchmark suites are built from TUDataset bioassay, molecule, and social-network collections (Dong et al., 2023). A notable recent application is dermoscopic imaging: images from HAM10000 are converted into graphs through segmentation, edge construction, and node feature extraction, and GLAD models are then applied in unsupervised, weakly supervised, and fully supervised regimes. In that setting, color descriptors give the best standalone performance, adding shape and texture consistently improves detection efficacy, and the best OCGTL configuration reaches $0.805$ AUC-ROC in the unsupervised regime, $0.872$ with 5% labeled anomalies, and $\mathcal{G}=(\mathcal{V},\mathcal{E})$0 with full supervision (Xu et al., 15 Aug 2025).

The supervision spectrum has also widened. Much of the classical literature assumes normal-only training, but newer work includes weakly supervised, fully supervised, and few-shot formulations. MA-GAD studies few-shot graph-level anomaly detection under 1-shot, 2-shot, 4-shot, and 8-shot settings, combining graph compression, meta-learning, and a deviation-based scoring module; it reports strong results on four biochemical datasets and argues that meta-learning is the more critical component relative to compression in ablations (Li et al., 9 Oct 2025). This suggests that GLAD is no longer a single regime, but a family of related tasks spanning one-class learning, binary graph classification, sparse-label adaptation, and even graph construction from non-graph raw data.

Current frontier themes are correspondingly broad: richer scoring functions, explicit spectral modeling, robustness to contamination and imbalance, interpretable anomaly evidence, and generalized OOD formulations. Across these directions, the central problem remains stable: to characterize normal graph populations precisely enough that whole graphs with deviant local structure, global organization, or spectral behavior can be identified reliably and, increasingly, explained.