This survey provides a comprehensive overview of OOD detection methods, organizing them into traditional and deep learning families before examining graph-aware GNN-based approaches and synthesizing the conceptual relationships between these paradigms through three working hypotheses: graph-aware transfer, structure-sensitive scoring, and hybrid objective design.
Abstract
Out-of-Distribution (OOD) detection, an important branch of Anomaly Detection (AD), aims to recognize inputs that fall outside the training distribution. Existing OOD methods based on statistical modeling, distance measures, one-class classification, reconstruction, or energy-based scoring provide strong foundations, but they can be difficult to apply directly to high-dimensional, unstructured, or relational data. Recent advances in graph-based deep learning, particularly Graph Neural Networks (GNNs), create new opportunities by modeling local and global dependencies among interdependent samples. This survey provides a comprehensive overview of OOD detection methods, organizing them into traditional and deep learning families before examining graph-aware GNN-based approaches. Beyond taxonomy, we synthesize the conceptual relationships between these paradigms through three working hypotheses: graph-aware transfer, structure-sensitive scoring, and hybrid objective design. This hypothesis-guided synthesis characterizes how established statistical, geometric, reconstruction-based, and energy-based principles can be transferred, reformulated, and combined within modern GNN frameworks. To make this perspective concrete, we include Graph Energy-based OOD Detection (GEO), a reference case study that combines energy-based scoring with one-class regularization in a GNN architecture and reports competitive but setting-dependent performance across representative benchmarks. Taken together, the survey clarifies the current landscape of OOD detection and highlights how integrating classical statistical reasoning with graph-based learning can inform future graph OOD research. To support reproducibility, the implementation of GEO is publicly available at https://github.com/longstnguyen/GEO.
Experiments on synthetic and real-world datasets show that ORExplainer consistently provides more robust explanations across diverse node-level OOD settings, outperforming existing baselines.
Geonhee Han, Heesoo Jung, Hyunju Kang et al.· Proceedings of the 32nd ACM...· 0 citations
Graph Neural Networks (GNN) facilitate effective prediction on graph data such as molecules, media networks and neural network blueprints. GNNs facilitate prediction through message passing techniques which define how information flows from a node to its neighbors. Due to the ubiquity of the graph data type, the development of newer and better GNNs has garnered much interest in the machine learning community. However, GNN evaluation and benchmarking is primarily driven by classification tasks. Thus, prospective GNN message passing layers are evaluated on their ability to outperform prior work in classification contexts. In contrast, GNNs are equally capable of performing scalar regression prediction, yet this class of problem is often overlooked when proposing new GNNs while the best classification GNNs are utilized in an a priori or off-the-shelf manner for regression problems. In response, this paper studies the efficacy of GNN layers in a slew of regression contexts from rank ordering, error minimization and insight extraction. Results show that deep convolutional GNNs, particularly GEN, are more effective at these tasks than attention-based GNNs, while other classical, theoretically-inspired GNNs remain competitive and efficient.
Keith G. Mills, Aedan J. DeFrates, Joong Ho Kim· 0 citations
This work proposes a novel Adaptive Dual-level Collaborative GNN associated with an adaptive dual-level collaborative mechanism, and shows that the ADC-GNN can inject the learned high-level information back into the node level, forming a closed-loop, bidirectional optimization process.
A novel framework, Generate and Filter graph learning for Graph Anomaly Detection (GFGAD), which generates a diverse set of synthetic anomalies with enriched feature and structural information to balance the data distribution and significantly outperforms state-of-the-art baselines.
Mengyu Li, Yonghao Liu, Ximing Li et al.· IEEE Transactions on Pattern...· 0 citations
In knowledge-intensive systems such as cybersecurity and financial risk control, graph data encode business rules, interaction constraints, and risk propagation paths through entities and relations. As a result, supervised graph anomaly detection (GAD) goes beyond conventional attribute outlier identification and becomes a problem of discovering anomalous structural behavior patterns. Under such conditions, the area under the precision–recall curve (AUPRC) better reflects practical risk-control objectives. From a mechanistic perspective, fixed neighborhood aggregation in end-to-end graph neural networks (GNNs) exhibits a low-pass bias in graph signal processing terms. At the same time, joint training tightly couples structural representation learning with a shallow classifier head of limited capacity, making high-curvature nonlinear decision boundaries difficult to learn under extreme imbalance and creating a capacity mismatch. However, upstream feature construction remains constrained by fixed low-pass operators, which creates spectral conflict because a strong classifier receives weak features. More importantly, structural camouflage is often common, whereas feature spectra may shift either left or right, leading to structural-spectral decoupling and spectral divergence. To address spectral conflict and capacity mismatch, we propose TAFH (Task-Aware Frequency Hybrid) for graph anomaly detection. TAFH explicitly decomposes multi-hop neighborhood propagation into low-frequency smooth components and high-frequency residual components, and then constructs an adaptive band-pass response through learnable hop weighting and node-level gated routing. To reduce objective drift under class imbalance, TAFH introduces a lightweight proxy head with a proxy objective, including a focal-loss variant, to inject hard-sample and minority-class preferences into representation learning. On the discrimination side, gradient boosted decision trees (XGBoost) provide high-capacity nonlinear interaction modeling through a decoupled boosting interface. Across five benchmark datasets (Reddit, YelpChi, Amazon, Tolokers, and Questions), TAFH consistently improves AUPRC over the strong baseline built from XGBoost and fixed neighborhood aggregation, with absolute gains from +1.12% to +6.35% and relative gains from 1.88% to 19.75%. These results show that supervised GAD can be cast as a coordinated process of structural knowledge encoding, adaptive frequency enhancement, and discriminative knowledge learning.