acceptodds
Withdrawn from ICLR 2027

Geometry-Aware Graph Anomaly Detection: Unified Hyperbolic Affinity Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised graph anomaly detection (UGAD) aims to identify abnormal nodes that deviate from normal patterns without labeled data and has been widely applied to fraud detection, cybersecurity, etc. However, node relationships in real-world graphs often exhibit hierarchical structure. Existing affinity-based graph anomaly detection methods mainly rely on Euclidean spaces to model node relationships, which limits their ability to capture relational differences in complex graph structures. Moreover, different relations are often modeled using different measurement spaces, making it difficult to jointly model multiple relations in complex graph structures. To address these issues, we propose a geometry-aware graph anomaly detection framework via unified hyperbolic affinity learning (UHAL-GAD). Specifically, we first map node representations into a shared Lorentz space and construct a unified hyperbolic affinity based on geodesic distances to learn more discriminative hyperbolic node representations. On this basis, the hyperbolic structural affinity learning module discriminates between normal and abnormal nodes using hyperbolic affinity, while the hyperbolic center affinity constraint module characterizes the hyperbolic center affinity discrepancy of each node with respect to the global Lorentz center. Furthermore, these modules jointly model multiple relations in a unified hyperbolic affinity measurement space. Extensive experiments on multiple graph datasets demonstrate the effectiveness and superior detection performance of UHAL-GAD.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Graph Anomaly Detection, Unsupervised Learning, Hyperbolic Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5oWEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents UHAL-GAD, an unsupervised graph anomaly detection method that maps node representations into Lorentz space. It combines neighborhood affinity learning with a global center-affinity constraint and derives anomaly scores from both relationships. Experiments on eight datasets report improvements over several existing detectors.

Critical strengths
  1. Combining local structural affinity and global center deviation within one geometric framework is a reasonable approach to unsupervised anomaly detection.

  2. The reported AUPRC results are strong across several datasets. The ablation results also indicate that both structural affinity and center constraints contribute to performance.

Critical weaknesses
  1. The need for hyperbolic geometry is not sufficiently established. The method uses standard Lorentz mappings and geodesic distances, followed by neighborhood and center-affinity objectives. Although a Euclidean ablation is provided, the paper does not establish whether the advantage comes from representing hierarchical structure or simply from a different affinity function and distance scale. More tightly matched geometric controls are needed.

  2. The global center constraint may suppress the anomalies it is intended to detect. Eq. (12) minimizes the center deviation of all nodes, including anomalous ones. The paper does not explain why this objective preserves abnormal representations rather than pulling them toward the normal center. This is especially important for clustered anomalies and multimodal normal distributions.

Questions for authors
  1. Can the authors isolate the effect of hyperbolic geometry from the affinity function and its scaling?

  2. Under what conditions does minimizing the global center loss preserve anomaly separation rather than causing representation collapse?

Reviewer_f9eiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes UHAL-GAD for unsupervised node-level graph anomaly detection. It maps node representations into a shared Lorentz space and uses geodesic-distance-based affinity for both local structural relationships and global center deviations. The framework consists of Hyperbolic Affinity Learning (HAL), Hyperbolic Structural Affinity Modeling (HSAM), and Hyperbolic Center Affinity Constraint (HCAC). HSAM distinguishes neighboring nodes from sampled non-neighbors, while HCAC measures deviations from a global Lorentz center. The two signals are combined for anomaly scoring. Results on eight graph datasets show the best reported AUPRC on all eight and the best AUROC on seven. Ablation and parameter sensitivity experiments are also included.

Critical strengths
  1. Combining local structural affinity with global center affinity addresses a limitation of neighborhood-based anomaly detection. Anomalous nodes forming strongly connected groups may still show high local consistency, making them difficult to distinguish from normal nodes. HCAC adds a global reference that may help in such cases. HSAM and HCAC also use the same hyperbolic affinity measure, so local and global relationships are modeled in a shared space.
  2. The results in Table 2 are favorable, especially for AUPRC, where UHAL-GAD outperforms all the reported baselines across eight datasets. Table 3 shows that removing HAL, HSAM, or HCAC lowers performance on the four evaluated datasets. The drops vary quite a bit, but each component has a measurable effect in these experiments.
Critical weaknesses
  1. The experiments do not yet isolate the benefit of hyperbolic geometry. The ablation provides some evidence, but the current description is not sufficient to show that the gains come from the hyperbolic formulation rather than from the combined local–global objective.
  2. The baseline comparison and unsupervised evaluation protocol are not sufficiently controlled. Closely related hyperbolic node-level methods are missing, and the paper does not clearly explain how hyperparameters and checkpoints are selected without anomaly labels, which raises concerns about fair comparison.
Questions for authors
  1. Can the authors provide a more controlled ablation that isolates the effect of hyperbolic geometry, e.g., by matching the Euclidean variant in architecture, parameter budget, and training settings and reporting it on all datasets?
  2. Can the authors justify the absence of hyperbolic node-level baselines, or add a comparable one? And can they clarify the label-free procedure for selecting β, γ, c, and checkpoints, including whether any anomaly labels were used?
Reviewer_uhTbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UHAL-GAD, an unsupervised graph anomaly detection framework that uses hyperbolic geometry to model node relationships. It combines local structural affinity between nodes and their neighbors with a global center constraint, and fuses the resulting signals to identify anomalous nodes. Experiments on eight graph datasets demonstrate promising detection performance.

Critical strengths
  1. The paper addresses an important problem in unsupervised graph anomaly detection by exploiting the hierarchical structure of graph data through hyperbolic geometry.
  2. The unified hyperbolic affinity provides a consistent way to model both local structural relationships and global center relationships.
  3. Experiments on graph datasets show promising detection performance, and ablation studies examine the contributions of the main components.
Critical weaknesses
  1. The proposed method combines hyperbolic geometry, local structural affinity, and global center constraints, some of which have been explored in prior graph anomaly detection methods. The novelty of this combination and its specific advantages over existing affinity-based methods remain insufficiently clear.
  2. The HCAC module minimizes the affinity discrepancy between all nodes and the global center during training, while this discrepancy is used as an anomaly signal at inference. It is unclear how this objective avoids pulling anomalous nodes toward the center and weakening their separation from normal nodes.
  3. HSAM models both positive and negative structural affinities during training, but the final anomaly score only uses positive affinity. The rationale for excluding negative affinity from anomaly scoring is unclear.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.