acceptodds
Under review as a conference paper at ICLR 2027

Unified Graph Anomaly Detection with Target-Context Evidence Reconstruction

Abstract

In graph-structured data, anomalous entities at different granularity levels (i.e., nodes, edges, and graphs) are often inherently interconnected. Motivated by such dependencies, unified graph anomaly detection (GAD) has recently emerged as a new research direction, aiming to detect multi-level anomalies within a single model. Although existing unified GAD approaches integrate multi-level detection tasks into a single framework, they typically convert detection objects at different granularities into a fixed subgraph format, overlooking the granularity-specific information required for anomaly detection at different levels. Moreover, existing methods still rely on level-specific detection heads or task-dependent references to support anomaly detection at different granularities, limiting knowledge integration throughout the end-to-end detection pipeline and preventing full task unification. To address these limitations, we propose Unified Target-context Reformulation And Contextual Evidence Reconstruction (UniTRACER) for unified GAD. To preserve granularity-specific information with a general task formulation, we introduce a target-context joint modeling paradigm and emphasize context importance based on the specific detection target. Moreover, to consolidate multi-level anomaly detection under a unified detection objective, we characterize the target-context relation and importance as shared detection evidence among levels, and then reframe unified GAD as an evidence reconstruction task. Together, these designs enable UniTRACER to retain granularity-specific anomaly cues while facilitating end-to-end knowledge sharing through a common task formulation and detection objective. Comprehensive experiments on 17 datasets demonstrate that UniTRACER significantly outperforms state-of-the-art granularity-specific and unified GAD methods.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Anomaly Detection, Graph Neural Network, Graph Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_svuZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniTRACER for unsupervised unified graph anomaly detection across node, edge, and graph levels. The method reformulates detection objects at different granularities as target-context pairs and estimates context importance conditioned on the target. The learned target-context relations and importance patterns are mapped into a shared four-dimensional evidence space. Anomaly detection is then formulated as an evidence reconstruction task using a shared autoencoder. Experiments on 17 datasets demonstrate consistent improvements over existing unified and granularity-specific baselines.

Critical strengths

1.The target-context formulation provides a clean and unified interface for multi-level anomaly detection. By explicitly separating the detection target from its surrounding context and conditioning the context representation on the target, the framework naturally adapts to the distinct information requirements of node, edge, and graph levels without relying on level-specific prediction heads.

2.The empirical evaluation is extensive. The paper evaluates the proposed method on 17 diverse datasets spanning citation, social, molecular, and pattern recognition graphs, comparing against 17 baselines. The comprehensive ablation studies and context importance visualizations provide strong support for the proposed methodology.

Critical weaknesses

1.The shared evidence vector in Eq. 6 consists of only four scalar values. The paper does not compare the autoencoder against other alternatives, making it unclear whether the reconstruction based scoring provides any benefit in this low dimensional evidence space. 2.There is a scale inconsistency within the shared evidence space. In Eq. 6, the first two components are negative cosine similarities , while the latter two are the mean and standard deviation of softmax probabilities, yielding extremely small values dependent on graph size. Directly concatenating these components without scale alignment causes the reconstruction loss to be entirely dominated by the cosine similarity terms. 3.The pseudo anomaly construction in Eq. 5 relies on randomly shuffling context graphs, which typically generates global structural noise rather than semantic or local topological anomalies. Relying on this may cause the model to merely distinguish normal structures from meaningless noise, and the effectiveness of this strategy requires further justification.

Questions for authors

1.What specific nonlinear modeling advantages does the autoencoder provide over simple statistical distance metrics in such a low dimensional space, and how is the necessity of this reconstruction based scoring empirically justified? 2.Given the scale discrepancy between the bounded cosine similarities and the extremely small probability based statistics in Eq. 6, what mechanism prevents the autoencoder reconstruction loss from being entirely dominated by the cosine similarity terms? 3.Random shuffling in Eq. 5 primarily generates structural noise rather than semantic anomalies. How does the model ensure that distinguishing normal structures from meaningless noise translates to the ability to detect subtle semantic graph anomalies?

2 comments
AuthorsOpenReview ↗

Response to Reviewer svuZ (Part 1)

We sincerely thank the reviewer for recognizing the clean and unified method design, the extensive empirical evaluation, the comprehensive ablation studies and context importance visualizations, and the clarity of the presentation.

Response to W1&Q1 - Necessity of Reconstruction-based Detector

We appreciate the reviewer’s question regarding the necessity of reconstruction-based anomaly detector. The key advantage of evidence reconstruction lies in modeling the complex relationships among evidence dimensions. Although the evidence vector contains only four scalars, these dimensions characterize different aspects of anomaly semantics and may exhibit nonlinear dependencies that hard to be captured by simple distance metrics. Unlike simple statistical distances that rely on predefined distributional assumptions, the autoencoder learns a data-adaptive reconstruction mapping to capture the underlying relationships among normal evidence patterns.

To further verify its effectiveness, we conduct additional ablation study by comparing the autoencoder-based detector against other alternative linear modelings, including a weighted sum approach that combines the three anomaly scores using learnable non-negative weights, and a linear fusion approach that employs a perceptron to integrate the scores. The results reported in terms of avg. AUROC are summarized below:

DetectorNode-levelEdge-levelGraph-level
AE (Ours)79.1680.3274.67
Weighted Sum68.2168.4066.68
Linear Fusion67.0968.3166.45

As the results demonstrate, both simpler baselines exhibit substantial performance degradation across node-, edge-, and graph-level anomaly detection. These findings indicate that, despite the low dimensionality of the evidence space, reconstruction-based anomaly scoring provides significant advantages over simpler linear modeling. This further highlights the importance of modeling the underlying relationships among evidence dimensions.

We have included a detailed analysis in the revised Appendix E.7 Additional Ablation Study on Detector.

Response to W2&Q2 - Contributions of Evidence Components

Thank you for this thoughtful observation. We agree that the evidence dimensions have different numerical scales, and appreciate the opportunity to further investigate how this affects evidence reconstruction. In our framework, the shared autoencoder jointly models the evidence dimensions through a reconstruction objective, where each dimension contributes according to its reconstruction residual. This allows the reconstruction loss to reflect the learned relationships among different evidence components, rather than being solely determined by their original numerical scales. To better understand this behavior, we further analyze the contribution of each evidence dimension to the overall reconstruction loss across different datasets in the revised Appendix E.8 Additional Analysis on Evidence Contribution.

From the Table 5 and 6, we observe distinct evidence contribution patterns across datasets. Specifically, target-specific evidence etgt\mathrm{e}_{\mathrm{tgt}} contributes substantially on citation networks (Cora, Citeseer, and ACM), where anomalies are often characterized by semantic inconsistencies between target nodes or edges and their neighborhoods. In contrast, Flickr exhibits a strong reliance on the average contextual importance mean⁡(I)\operatorname{mean}(\mathbf{I}). For molecular and biochemical graphs such as DHFR, PTC-MR, and PROTEINS, importance-based evidences contribute substantially, reflecting the relevance of intra-graph context-selection patterns to graph-level abnormalities.

Overall, these results demonstrate that the reconstruction loss is not systematically dominated by evidence dimensions with larger numerical scales. Instead, different evidence components contribute differently depending on the underlying anomaly characteristics, highlighting the benefit of jointly modeling complementary anomaly evidence within a shared reconstruction framework.

AuthorsOpenReview ↗

Response to Reviewer svuZ (Part 2)

Response to W3&Q3 - Context Corruption

We thank the reviewer for this valuable comment. In fact, permutation-based corruption has been widely adopted as an effective negative sample construction strategy in previous studies [1, 2]. By disrupting the inherent correspondence between graph entities and their original neighborhood contexts, shuffling creates contextually inconsistent samples that effectively simulate deviations from normal relationships, thereby providing meaningful supervisory signals for learning discriminative information.

To further validate the effectiveness of this strategy, we conduct additional experiments comparing our permutation-based corruption (Shuffle) with two alternative strategies designed to introduce semantic anomalies (Feature Noise) and topological anoamlies (Edge Rewiring). The average AUROC (%) results are summarized below:

MethodNode-levelEdge-levelGraph-level
Shuffle (Ours)78.4279.0777.32
Feature Noise73.1975.9876.34
Edge Rewiring76.4277.4674.26

As shown, our permutation-based corruption consistently outperforms both alternatives across node-, edge-, and graph-level anomaly detection. These results demonstrate the effectiveness of shuffling in generating meaningful contextual anomalies and learning discriminative representations capable. Moreover, shuffling introduces minimal computational overhead, preserving the efficiency of our unified anomaly detection framework.

We have included these results and detailed discussions in the revised Appendix E.6, Additional Ablation Study on Corruption Strategies.

[1] Park, N., Rossi, R., Koh, E., Burhanuddin, I. A., Kim, S., Du, F., ... & Faloutsos, C. (2022, April). Cgc: Contrastive graph clustering forcommunity detection and tracking. In Proceedings of the ACM web conference 2022 (pp. 1115-1126).

[2] Long, Y., Ang, K. S., Li, M., Chong, K. L. K., Sethi, R., Zhong, C., ... & Chen, J. (2023). Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST. Nature communications, 14(1), 1155.

Reviewer_RxkbOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work introduces UniTRACER, an unsupervised method for unified GAD that detects multi-level anomalies within a single model. It first identifies two challenges in unified GAD at the granularity unification level and detection level, highlighting the technical difficulty of unifying different levels of GAD tasks with unique requirements. To address them, the authors then introduce a target-context joint modeling paradigm to allow the multi-level tasks to be unified while keeping their unified requirements. Moreover, the authors further employ evidence reconstruction as a task to unify the semantics of multi-level tasks, thereby enabling effective sharing of knowledge for UniGAD. The comprehensive evaluation on 17 datasets demonstrate UniTRACER's effectiveness.

Critical strengths
  1. This paper is well-motivated. It points out the key dilemma of how to unify tasks with different requirements in GAD. From a context perspective, different granularities of tasks prefer different contexts, making unifying tasks challenging. While from a task semantic perspective, it also requires detecting anomalies at different levels with different semantics using a single detector to achieve knowledge sharing. These two challenges lie at the heart of the difficulty in designing a UniGAD method, and I believe identifying them can help this field progress beyond simply stacking multiple detectors into one.

  2. The method is elegantly designed to address the aforementioned challenges. It introduces a context selection mechanism to obtain task-related context, thereby enabling task unification while ensuring granularity-specific information can be highlighted. Compared to heuristic-based designs such as removing heterophily edges, this design can better capture long-range dependencies. Moreover, it further extracts evidence from the representations to align multi-level detection tasks into a unified space for detection. This enables the model to effectively share knowledge between tasks without employing a dedicated detection head for each level. Notably, the effectiveness of the designed methods is comprehensively evaluated in a large number of experiments.

  3. The paper is well written and easy to follow. The formulation for each component is clearly introduced. The figures are clear to to help demonstrate the design of the proposed method.

Critical weaknesses
  1. While the paper provides the hyperparameter search space, including a more detailed sensitivity analysis would be helpful for further study. Notably, this paper uses a fixed context neighborhood size across experiments, which should be analyzed in more detail, as it is a key component for constructing context.

  2. The interpretability analysis is mainly qualitative, while no quantitative results are provided. The authors should consider including additional quantitative results to evaluate its effectiveness.

  3. There are a few designs of the proposed method need further explanation. Details please see the "Questions For Authors" session.

Questions for authors
  1. Can the authors explain why using the three evidences are helpful for anomaly detection?

  2. Regarding the anomaly detectors, despite the effectiveness demonstrated in ablation, I am wondering why an autoencoder is employed as the anomaly detector, rather than direct using a mlp based detector and train with contrastive learning objective?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Rxkb

We sincerely thank the reviewer for recognizing the well-motivated challenges, the elegant method design, the comprehensive experimental evaluation, and the clarity of the presentation.

Response to W1 - Sensitivity Analysis

We appreciate the reviewer’s suggestion to provide a more comprehensive parameter sensitivity analysis. Accordingly, we have conducted additional experiments to investigate the sensitivity of UniTRACER to four key hyperparameters: K, alpha, beta, and gamma. The results and detailed analyses are presented in the revised Appendix E.9, Additional Parameter Study.

Regarding the context neighborhood size ((K)), our results consistently show that (K=2) achieves the best performance across datasets. This suggests that a 2-hop neighborhood provides sufficient local structural and contextual information for effective anomaly detection. Increasing (K) beyond 2 introduces additional computational overhead without yielding consistent performance improvements. Therefore, we adopt a fixed context neighborhood size of (K=2) across all experiments to achieve a favorable balance between detection effectiveness and computational efficiency.

Response to W2 - Quantitative Interpretability

We appreciate the reviewer’s valuable suggestion. We would like to highlight that Table 1 already provides quantitative evidence for anomaly interpretability through the BM-MN and BM-MS datasets. Specifically, these two benchmarks contain ground-truth explanatory motifs, where the nodes and edges belonging to these motifs serve as fine-grained anomaly annotations. Therefore, node- and edge-level detection performance on these datasets quantitatively reflects the model's ability to identify the underlying anomalous substructures responsible for graph-level abnormalities.

As shown in Table 1, UniTRACER achieves strong performance on both datasets, with node-level AUROC of 81.67% and 87.27%, and edge-level AUROC of 83.22% and 87.60%, respectively. These results demonstrate its effectiveness in identifying anomaly-related nodes and edges.

Response to Q1 - Necessary of three evidences

We appreciate the reviewer’s question. The three evidences capture complementary aspects of target-context relationships. Specifically, holistic relation evidence measures the overall consistency between the target and its context, while target-specific evidence focuses on their consistency after emphasizing target-relevant contextual information. Importance pattern evidence further characterizes how contextual relevance is distributed, capturing information beyond the aggregated representations. Together, they provide a more comprehensive characterization of anomaly patterns across different granularities. As demonstrated in Fig. 3(b) and Appendix E.3, combining all three evidences consistently outperforms other variants.

Response to Q2 - Necessary of Autoencoder

We thank the reviewer for this valuable suggestion. We employ an autoencoder because evidence reconstruction provides a natural unsupervised detection objective that jointly models the relationships among different evidence dimensions, without requiring explicitly constructed negative samples.

To further evaluate the necessity of evidence reconstruction, we additionally compare our autoencoder with two simpler detector alternatives: a learnable nonnegative weighted sum and a linear classifier. The average AUROC results are summarized below:

DetectorNode-levelEdge-levelGraph-level
AE (Ours)79.1680.3274.67
Weighted Sum68.2168.4066.68
Linear Fusion67.0968.3166.45

As shown, the reconstruction-based detector consistently outperforms both alternatives across all three anomaly granularities. These results suggest that learning the joint structure of evidence through reconstruction is more effective than directly combining evidence scores, further supporting our choice of the autoencoder for unified GAD.

We have included a detailed analysis in the revised Appendix E.7 Additional Ablation Study on Detector.

Reviewer_xVJROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work studies unified graph anomaly detection (GAD), which aims to jointly identify anomalies at the node, edge, and graph levels. The authors identify two key limitations of existing unified GAD frameworks: (i) they homogenize multi-granularity objects into a fixed subgraph representation, which obscures the granularity-specific information required by different detection tasks, and (ii) they rely on task-dependent detection heads, whose fragmented design hinders end-to-end knowledge integration across granularities. To address these issues, the paper proposes UniTRACER, built upon a target-context joint modeling paradigm. Extensive experiments on 17 datasets against 17 baselines show that UniTRACER achieves state-of-the-art AUROC in 16 out of 18 dataset-granularity combinations, with computational complexity comparable to that of a vanilla GNN.

Critical strengths
  1. The paper provides a diagnosis of existing unified GAD pipelines. Distinguishing the loss of granularity-specific information caused by fixed-format homogenization (RQ1) from the fragmentation caused by task-specific detection heads (RQ2) is conceptually valuable.
  2. Reformulating unified detection as an evidence reconstruction task with a shared autoencoder is a clean way to consolidate node-, edge-, and graph-level detection into a single objective.
Critical weaknesses
  1. A comparison with a simple baseline that fuses these scores without an autoencoder (e.g., a weighted sum or a concatenation followed by a linear classifier) is missing, so it remains unclear how much the evidence reconstruction formulation itself contributes.
  2. The paper only states that the coefficients are selected by grid search; a sensitivity analysis of alpha, beta, and gamma and a discussion of their interactions across datasets are absent.
  3. The pseudo-anomalies used by the discriminability loss are generated by randomly shuffling the context. Such synthetic perturbations may differ substantially from real anomaly mechanisms, such as feature camouflage or motif-level structural defects. The paper neither discusses the failure cases of this design nor compares alternative context corruption strategies, such as feature noise or edge rewiring.
  4. In the graph-level setting, the entire graph serves as both the target and the context. The paper does not explain how context importance can be meaningfully computed when the context is identical to the target, nor what anomaly semantics the importance pattern evidence captures in this degenerate case.
Questions for authors
  1. Can you add an experiment comparing the shared evidence autoencoder with simple score-level fusion, such as a weighted sum or a linear fusion of the three evidence values, in order to isolate the contribution of evidence reconstruction?
  2. How sensitive is the performance to the loss coefficients alpha, beta, and gamma and to the context hop size K? Are the optimal values stable across datasets, and have you considered an adaptive K?
  3. For graph-level detection where the target and the context are the same graph, how is the context importance score computed and interpreted, and what kind of graph-level anomalies does the importance pattern evidence identify?
  4. Could you provide a real-world case study in which a single anomaly manifests across granularities, for example a fraudulent node inducing anomalous edges, showing how the target-context evidence captures such cross-level dependencies?
2 comments
AuthorsOpenReview ↗

Response to Reviewer xVJR

We sincerely thank the reviewer for recognizing the value and clarity of UniTRACER, its state-of-the-art performance, efficiency, and the comprehensive experimental evaluation, as well as for the constructive suggestions on additional experiments and further clarifications.

Response to W1&Q1 - Anomaly Detector Baselines

We appreciate the reviewer’s question regarding including additional simpler baselines without an autoencoder for comparison. We agree that additional simpler anomaly detector baselines can be included to evaluate how much the reconstruction contribute to UniTRACER.

Accordingly, we incorporate the weighted sum and linear classfier into our comparison. The results are reported below in avg. AUROC:

DetectorNode-levelEdge-levelGraph-level
AE (Ours)79.1680.3274.67
Weighted Sum68.2168.4066.68
Linear Fusion67.0968.3166.45

As the results demonstrate, both simpler baseline's performance drops across node-, edge-, and graph-level anomaly detection. These findings indicate that simply combining the three evidence scores is insufficient for effective unified GAD, highlighting the substantial contribution of our reconstruction-based formulation beyond score-level fusion.

We have included a detailed analysis in the revised Appendix E.7 Additional Ablation Study on Detector.

Response to W2&Q2 - Sensitivity Analysis

We appreciate the reviewer’s suggestion to include additional experiments and analyses on parameter sensitivity. Accordingly, we have conducted a detailed sensitivity analysis of K, alpha, beta, and gamma, which is now included in the revised Appendix E.9 Additional Parameter Study.

Our results show that the optimal values of alpha and the hop size K remain consistent across datasets (alpha = 1, K = 2), whereas the optimal values of beta and gamma vary across datasets. In particular, increasing K beyond 2 hops introduces additional computational overhead without yielding further stable performance improvements. These findings suggest that a fixed hop size of (K = 2) is sufficient for UniTRACER. Nevertheless, we believe that exploring adaptive hop selection represents a promising direction for future research.

Response to W3 - Alternative Context Corruption Strategies

We appreciate the reviewer’s suggestion to investigate alternative context corruption strategies for computing the discriminability loss. Accordingly, we conduct additional experiments comparing our feature shuffling strategy with two alternatives: feature noise injection and edge rewiring. The results, reported in terms of average AUROC (%), are summarized below:

MethodNode-levelEdge-levelGraph-level
Shuffle (Ours)78.4279.0777.32
Feature Noise73.1975.9876.34
Edge Rewiring76.4277.4674.26

As the results demonstrate, despite its simplicity, our feature shuffling strategy consistently outperforms both alternatives across node-, edge-, and graph-level anomaly detection. These findings highlight the effectiveness of feature shuffling in proxying real anomaly mechanisms for Unified GAD.

We have included a detailed analysis in the revised Appendix E.6 Additional Ablation Study on Corruption Strategies.

AuthorsOpenReview ↗

Response to Reviewer xVJR (second part)

Response to W4 & Q3 - Interpretation fo Graph-level Detection

We thank the reviewer for this insightful question. For graph-level anomaly detection, although the target and context correspond to the same graph, they play different semantic roles during importance estimation.

Specifically, in graph-level anomaly detection, context importance can be interpreted as a global-to-local relevance score, which highlights regions most relevant to anomaly detection through a self-attention-like mechanism. The target representation captures the holistic semantics of the graph, serving as a global semantic reference. This representation then guides the importance estimation of individual nodes, highlighting their relative relevance under the graph-specific semantic context. Therefore, the identity between the target and context does not make the importance estimation degenerate, as the importance reflects the relevance of local components to the global graph semantics.

This is particularly beneficial for graph-level anomaly detection, where anomalies often arise from distinctive local structural patterns (e.g., meanful functional groups in molecular graphs) that are highly relevant to the task domain. By leveraging graph-specific semantic information, our target-context mechanism can effectively highlight these informative local patterns, enabling UniTRACER to better capture anomaly-relevant information.

Our ablation results further support this interpretation. As shown in Fig. 3(b), removing importance pattern evidence leads to a substantial performance drop at the graph level. Moreover, Appendix E.3 shows that this evidence alone achieves strong graph-level detection performance, highlighting its effectiveness for graph-level anomaly detection.

Response to Q4 - Real-world Case Study

We appreciate the reviewer’s suggestion to incorporate additional real-world case study. We have added a real-world case study on the DHFR dataset in Appendix E.10 Additional Case Study on DHFR dataset, illustrating how UniTRACER captures cross-level dependencies within a single anomalous molecular graph.

Specifically, in this dataset, molecules exhibiting DHFR inhibitory activity are considered anomalous. As shown in Figure 15, the graph-level importance scores highlight two key functional groups (i.e. motifs in the graph) within the molecule. On the right-hand side, the -NH2 groups in 2,4-diaminopyrimidine are highlighted. On the left-hand side, the carbon skeleton (C for carbon atoms and H for hydrogen atoms) in the dichloro-substituted aromatic group also receives high importance scores. These groups are associated with DHFR inhibitory activity through hydrogen bonding and hydrophobic interactions, respectively, as supported by existing chemical studies [1,2]. This demonstrates that our target-context evidence can identify chemically meaningful local structures relevant to the graph-level anomaly.

Furthermore, the node- and edge-level anomaly scores of the same molecule also highlight these functional groups. Importantly, these anomalous nodes and edges collectively constitute the functional groups associated with the molecule's inhibitory activity. This correspondence illustrates how local structural anomalies at the node and edge levels relate to the overall graph-level anomaly, and our UniTRACER can capture meanful cross-level anomaly dependencies.

[1] Spowage, B. M., Bruce, C. L., & Hirst, J. D. (2009). Interpretable correlation descriptors for quantitative structure-activity relationships. Journal of cheminformatics, 1(1), 22.

[2] Nelson, R. G., & Rosowsky, A. (2001). Dicyclic and tricyclic diaminopyrimidine derivatives as potent inhibitors of Cryptosporidium parvum dihydrofolate reductase: structure-activity and structure-selectivity correlations. Antimicrobial agents and chemotherapy, 45(12), 3293-3303.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.