acceptodds
Under review as a conference paper at ICLR 2027

Generalist Graph Anomaly Detection via a Spectral Vocabulary

Abstract

Graph anomaly detection (GAD) methods are typically trained on individual graphs, limiting their ability to transfer anomaly knowledge across domains. Developing a generalist GAD model therefore requires representations that support transfer to unseen graphs. A fundamental challenge is the lack of a shared representation space across graphs that effectively leverages diverse graph data while generalizing across domains. To address this challenge, we introduce SV-GAD, a generalist GAD framework built on a shared "graph vocabulary". Our key insight is that spectral-energy statistics characterize how node signals vary with respect to graph structure through common measurements, providing a domain-agnostic representation across graphs. SV-GAD learns transferable anomaly patterns in the spectral space and leverages a mixture-of-experts architecture, where graph filters serve as specialized experts for different spectral behaviors. Extensive experiments across diverse GAD benchmarks demonstrate that SV-GAD effectively transfers anomaly knowledge to unseen graphs and consistently outperforms existing baselines in a zero-shot setting.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Anomaly Detection, Generalist Graph Anomaly Detection, Spectral Energy, Zero-Shot

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jjLeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SV-GAD for zero-shot graph anomaly detection across heterogeneous domains. It constructs a shared spectral vocabulary from local graph-signal energy and uses a mixture of spectral experts to learn transferable anomaly representations. Experiments on eleven target datasets show strong results, particularly on citation and transaction graphs.

Critical strengths
  1. The spectral representation offers an interesting way to address feature mismatch across graphs. The underlying spectral-energy analysis provides useful motivation.
  2. The zero-shot evaluation is extensive. SV-GAD achieves the best AUROC and AUPRC on eight of eleven target datasets, with particularly strong results on several citation graphs.
Critical weaknesses
  1. There is an inconsistency between the theoretical bound and the implemented vocabulary. Proposition 4.2 bounds the column sum of the spectral vocabulary by 4, but Eq. (7) multiplies the weighted local energy by the number of nodes. The stated bound does not directly hold for this rescaled representation. Since boundedness is used to justify cross-graph comparability, this needs correction.
  2. The independent contribution of the spectral vocabulary is not fully isolated. The final model concatenates spectral descriptors with projected features, and the ablations mainly examine the other architectural components. A controlled comparison using spectral features alone, projected features alone, and their combination would better establish the source of the transfer gains.
Questions for authors
  1. Does Proposition 4.2 apply to the actual representation in Eq. (7), including its factor of N?
  2. How much of the final improvement comes from the spectral vocabulary itself rather than the aligned features and attention backbone?
Reviewer_YE4iOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces SV-GAD for zero-shot generalist graph anomaly detection. It combines a shared spectral vocabulary based on normalized local spectral-energy statistics with projected node features, neighborhood attention, and learnable spectral experts. Trained on labeled source graphs, it transfers to unseen graphs without target labels or fine-tuning, achieving the best AUROC and AUPRC on 8 of 11 target datasets.

Critical strengths

1.The spectral vocabulary addresses semantic differences that feature-dimension alignment alone cannot resolve, providing an interpretable basis for cross-graph comparison.

2.The spectral experts capture anomalies across frequency bands, while node-specific routing and neighborhood attention combine these patterns, supported by ablations.

3.Experiments span diverse domains and recent generalist baselines, with particularly strong gains on citation graphs. Evaluation on the 5.78-million-node T-Social dataset also demonstrates scalability.

Critical weaknesses

1.I am not convinced that the proposed method adequately resolves feature heterogeneity. It still relies on independently fitted PCA/SVD projections, followed by fixed operations such as spectral-energy normalization and channel sorting. Stronger evidence is needed to show that the spectral vocabulary overcomes these limitations rather than primarily benefiting from dimensionality reduction and preprocessing.

2.The mixture-of-experts and multi-head attention components are widely used architectural techniques, and the paper does not clearly establish a distinctive methodological advance in their combination. The main novelty therefore appears to lie in the spectral vocabulary as a feature extraction scheme. This narrows the scope of the contribution and makes its significance depend primarily on whether the proposed features provide clear advantages over simpler structural and spectral descriptors.

3.Could the authors clarify the following aspects of Eq. (11)?

  • Motivation: Why should routing probabilities align with spectral-energy proportions? A band containing more energy does not necessarily provide more useful information for anomaly detection.
  • Warm-up configuration: What values are used for the regularization coefficient γ\gamma and the number of warm-up epochs? Are they fixed across all datasets or tuned separately? If tuned separately, how are they selected without target-domain labels under the strict zero-shot setting?

4.The experimental setup states that hyperparameters are selected through random search on source graphs, but does not specify the selected values used in the main experiments. For example, what is the default channel dimension CC? Please provide the search ranges and final settings for key hyperparameters like regularization coefficient γ\gamma, and warm-up duration. Please also clarify whether the same configuration is used across all target datasets and how configurations are selected using only source data.

Questions for authors

Please see Weaknesses

Reviewer_dDHqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SV-GAD for zero-shot node-level anomaly detection on unseen graphs. It constructs shared node representations using spectral-energy statistics and combines graph attention with a mixture of spectral experts. Experiments report improvements across multiple datasets.

Critical strengths

Using local feature disagreement to construct transferable graph representations is a promising idea. The reported results are competitive, particularly on citation networks, and successful evaluation on the large T-Social dataset is encouraging.

Critical weaknesses

The theoretical formulation is inconsistent with the implemented representation. Equation (1) does not hold as stated: on a three-node path, considering the center node with signal x=(0,1,0)⊤x=(0,1,0)^\top, the normalized-Laplacian Rayleigh quotient is 11, whereas ηi/δi=2\eta_i/\delta_i=2. Furthermore, Proposition 4.2 establishes a bound below 44 for Eˉi\bar E_i, while Equation (7) uses NEˉiN\bar E_i, which need not satisfy that bound. These inconsistencies weaken the theoretical justification for the vocabulary’s properties and cross-graph comparability.

A key control for the central contribution is missing. The ablations remove the projected features ZZ, but do not remove the spectral vocabulary SS while retaining the same backbone. Consequently, the vocabulary’s contribution to the overall improvement remains unclear. Feature expansion, routing-target construction, and baseline adaptations are also insufficiently specified, and the main results lack variability estimates across random seeds.

Questions for authors

How should Equation (1) be corrected, and which theoretical guarantees apply to the representation actually used in Equation (7)? Can you compare Z-only, S-only, and [S∥Z][S\Vert Z] inputs under the same backbone, with variability estimates and the implementation details needed to reproduce the results?

Comments on the paper

1 comment

Dear Reviewers,

This is a friendly reminder to check the authors' responses to your reviews and leave a follow-up comment if appropriate. Your engagement in the discussion is greatly appreciated.

Thank you!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.