acceptodds
Under review as a conference paper at ICLR 2027

From Single- to Multi-relation: Relation Inconsistency Modeling for Unsupervised Graph Anomaly Detection

Abstract

Graph anomaly detection (GAD) in this work aims to identify anomalous nodes whose behavioral patterns deviate from those of the majority of normal nodes. However, most existing GAD methods are developed for single-relation graphs and cannot directly capture the diverse relational semantics prevalent in real-world systems. Current multi-relation approaches, in turn, largely overlook *cross-relation behavioral inconsistency* as an informative signal of anomalies. To bridge this gap, we identify two complementary forms of relation inconsistency: *explicit inconsistency*, manifested as structural connectivity shifts across relations, and *implicit inconsistency*, revealed by divergent reconstruction patterns. Based on this insight, we propose RelIn-GAD, an architecture-agnostic, plug-and-play module that extends unsupervised reconstruction-based single-relation detectors to multi-relation settings. RelIn-GAD employs global-to-local relation weighting to adaptively aggregate relation-wise reconstruction evidence, and a dual-inconsistency mechanism that distinguishes anomaly-related disparities from benign relational diversity. The resulting inconsistency signal further adapts the node-wise training objective, while explicit structural inconsistency provides a gated correction to reconstruction-based anomaly scoring. Extensive experiments on four real-world datasets and seven backbones demonstrate the broad effectiveness of RelIn-GAD, yielding average relative improvements of up to 36.29% in AUROC and 108.89% in AUPRC, while achieving strong performance against eighteen state-of-the-art baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
anomaly detection, multi-relation graph, cross-relation inconsistency

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_KY2aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates unsupervised multi-relational graph anomaly detection and proposes an architecture-agnostic, plug-and-play method RelIn-GAD. RelIn-GAD iteratively aggregates relation-wise reconstruction evidence by global-to-local relation weighting and distinguishes anomaly-related disparities via a dual-inconsistency mechanism. Extensive experiments show the effectiveness and the generalization of RelIn-GAD on multiple datasets.

Critical strengths

(1) The proposed method RelIn-GAD is an plug-and-play module, which can be easily integrated into other baselines and achieves consistent improvements.

(2) The motivations of most modules are reasonable, especially the inconsistency measurement and mathematical formulation of the modules.

(3) The paper is well-writen and easy to follow.

Critical weaknesses

(1) The ablation study (Table 3) cannot illustrate the effectiveness of each module. On most of dataset in Table 3 (except Amazon), the removal of each module has no significant performance drop.

(2) In Table 1 and 2, I discover AUROC of many baselines on several datasets are below 50%. A random baseline should be 50%. Is there any experiment fault in these two tables?

Questions for authors

See weaknesses.

Other questions:

(1) The recent baselines UMGAD and SFGA are both the multi-relational GAD baselines. What are the technical differences between these baselines and the proposed method?

(2) In my opinion, multi-relational graph is one category of the heterogenous graph with just one kind of nodes. Is there any heterogenous GAD method which can serve as the baselines?

4 comments

Thank you for your constructive feedback to help improve our paper. Below, we address your concerns point by point, which has been reflected in the revised manuscript. Please kindly let us know if our response addresses your concerns.

W1. Clarification of ablation results

R1. We thank the reviewer for this observation. The contributions of individual components depend on the characteristics of cross-relation inconsistency in each dataset, rather than producing uniform gains in all settings. For example, removing explicit or implicit inconsistency leads to noticeable performance degradation on Amazon, whereas the effects are smaller on some other datasets. This is consistent with our framework's objective: exploiting informative cross-relation disparities when they are present, rather than assuming all relational differences are equally useful for anomaly detection. In particular, on IEEE-CIS, only 0.28% of nodes belong to the implicit-inconsistent category, while 74.91% are relation-consistent. Consequently, the implicit mechanism has limited influence on the overall anomaly ranking, which helps explain why its removal produces close AUROC/AUPRC values. The additional taxonomy ablation also compares degree-only, reconstruction-only, ungated dual-signal, and random-gating alternatives under matched configurations.

We have revised the ablation discussion in Sec. 4.5 (Line 451) and added further dataset-dependent analysis in Appendix H (Line 1080) of the revised manuscript.

W2. Explanation of AUROC results below 50%

R2. The reported AUROC values below 50% are valid outcomes under our evaluation protocol and do not, by themselves, indicate experimental errors. AUROC measures the probability that a randomly selected anomalous node receives a higher anomaly score than a randomly selected normal node. Therefore, while random ranking has an expected AUROC of 50%, a detector can legitimately achieve a lower value when its scoring assumptions do not align with the anomaly patterns in a particular dataset. For single-relation GAD baselines, we merge edges from different relations into a homogeneous graph, following the evaluation strategy adopted by SFGA [2]. This transformation removes relation-specific semantics and may weaken the effectiveness of detectors originally designed for homogeneous graphs. Consequently, some baselines produce anomaly rankings that are less informative than random ranking on particular datasets. We report these results as obtained, without artificially reversing anomaly scores or adjusting AUROC values to exceed 50%.

[2] Huiliang Zhai, Xiangyi Teng, Jing Liu, SFGA: Similarity-Constrained Fusion Learning for Unsupervised Anomaly Detection in Multiplex Graphs. AAAI 2026.

Q1. Technical differences from UMGAD and SFGA

R3. We kindly refer the reviewer to the detailed methodological comparison already provided in Appendix C (Line 830) of our original submission. Although all three methods address multi-relation graph anomaly detection, their core mechanisms differ. UMGAD employs graph-masked autoencoders with augmented-view contrastive learning for anomaly detection, while SFGA focuses on cross-relation attention and similarity-constrained representation learning. In contrast, RelIn-GAD explicitly distinguishes structural and reconstruction-level inconsistencies, incorporating both into node-adaptive training while using gated structural inconsistency for inference correction. Furthermore, unlike these standalone architectures, RelIn-GAD provides a plug-and-play enhancement for existing reconstruction-based detectors. We have further added a brief clarification of these technical distinctions in Sec. 4.4 (Line 428) of the revised manuscript.

Q2. Additional heterogeneous graph anomaly detection baselines

R4. We thank the reviewer for this valuable suggestion. Multi-relation graphs can be viewed as a special case of heterogeneous graphs with a shared node set and multiple relation types. However, relatively few studies specifically address unsupervised node-level anomaly detection in heterogeneous graphs, and the official implementation of a recent unsupervised heterogeneous graph anomaly detection work, HINAno [3], is not publicly available. Therefore, we additionally adapt two recent self-supervised heterogeneous graph learning methods with publicly available implementations, MUG [4] and MVMAE [5], for comparison. Both methods are adapted to our multi-relation setting by preserving relation-specific graph views and introducing reconstruction-based node-level anomaly scoring.

Table I: Comparison with adapted heterogeneous graph methods (AUROC / AUPRC, %).

MethodAmazonYelpChiS-FFSDIEEE-CIS
MUG74.11 / 19.4351.15 / 15.9456.97 / 25.0454.26 / 15.43
MVMAE51.87 / 6.8449.26 / 13.9231.90 / 13.4857.28 / 10.32
RelIn-GAD88.41 / 36.6662.69 / 24.1564.75 / 29.4969.33 / 18.91

As shown in Table I, our evaluation extends to recent heterogeneous graph representation learning methods adapted for unsupervised anomaly detection. They reflect our added reconstruction-based scoring adaptations, rather than anomaly detection results originally reported by those papers. Unlike these adapted methods, RelIn-GAD models both structural and reconstruction-level inconsistencies during training, while using gated structural inconsistency as complementary evidence for inference scoring. We have referenced the supplementary comparison in Sec. 4.1 (Line 338) and added the adapted heterogeneous graph learning baselines and experimental results in Appendix D (Line 879) of the revised manuscript.

[3] Xiangjie Kong, Siyue Shuai, Hui Wang, Guojiang Shen, Feng Xia, Dual-View Anomaly Detection in Heterogeneous Information Networks with Hierarchical Neighborhood Fusion. ACM TKDD 2025.

[4] Lianze Shan, Jitao Zhao, Dongxiao He, Yongqi Huang, Zhiyong Feng, Weixiong Zhang, MUG: Meta-path-aware Universal Heterogeneous Graph Pre-Training. AAAI 2026.

[5] Hao Geng, Deqing Wang, Lihao Ge, Chenguang Du, Fuzhen Zhuang, Bowen Yang, Efficient and Scalable Multiview Masked Autoencoder for Heterogeneous Graph Self-Supervised Learning. IEEE TNNLS 2026.

Thank the authors for their response, which has addressed my concern to some extent. However, the innovations proposed in this paper are not particularly compelling to me, so I will maintain my original score.

Dear Reviewer KY2a,

We sincerely thank you for your time and thoughtful consideration of our response. We appreciate your acknowledgment that our clarifications have addressed some of your concerns.

While we respect your assessment, we would like to briefly reiterate two key aspects of our novelty and contributions:

  1. Relation Inconsistency Modeling: Unlike UMGAD and SFGA, which primarily focus on contrastive learning and cross-relation representation fusion, respectively, RelIn-GAD explicitly models two complementary forms of cross-relation inconsistency (structural and reconstruction-based), leveraging them for node-adaptive training and gated anomaly scoring.

  2. Plug-and-Play Framework: Rather than designing a standalone multi-relation detector, RelIn-GAD provides an architecture-agnostic framework that extends existing reconstruction-based single-relation GAD detectors to multi-relation settings without backbone redesign. Its effectiveness is further supported by experiments across seven backbones and additional comparisons with heterogeneous graph methods.

We would be grateful for any further guidance on your remaining reservations. In particular, it would be helpful to know whether we could better clarify the methodological distinctions from prior work or the significance of the plug-and-play framework, or whether there are other aspects that remain insufficiently addressed.

Thank you again for your valuable feedback and consideration.

Best regards, The Authors

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

The paper proposes RelIn-GAD, a module extending reconstruction-based graph anomaly detectors to multiple relations. It combines adaptive relation weighting, structural and reconstruction inconsistencies, and inconsistency-aware training. Experiments cover four datasets and seven backbones.

Critical strengths

Addresses a relevant problem with a practical, reusable framework.

Evaluates multiple backbones and reports AUROC/AUPRC over five runs.

Provides component ablations, sensitivity analyses, case studies, and runtime measurements. Improvements on Amazon are substantial.

Critical weaknesses

Vanilla backbones collapse relations into one graph, whereas RelIn-GAD preserves relation-specific information. Appendix D includes uniform weighting, but a matched multi-relation baseline removing all inconsistency mechanisms is still needed. RelIn-GAD receives grid search while other baselines largely use defaults; the label usage and validation protocol for model selection are unclear.

Section 4.2 compares disagreement among small predicted anomaly sets with agreement among their much larger complements. This asymmetry can arise from set sizes alone and does not establish that disagreement distinguishes true anomalies. A size-matched random reference and label-conditioned analysis would strengthen the motivation.

Equation (6) increases reconstruction pressure on highly inconsistent nodes, potentially improving reconstruction of the anomalies being detected. Appendix E establishes the gradient multiplier but does not explain why this improves anomaly separation.

In Table 3, removing either inconsistency signal leaves IEEE-CIS results unchanged. Table 2 also shows substantially higher Amazon AUPRC for FreeGAD (69.33 versus 35.58). These results warrant more qualified claims about overall superiority.

Questions for authors
  1. How are hyperparameters and checkpoints selected without evaluation-label leakage?

  2. How much improvement remains over a matched multi-relation backbone with uniform aggregation and no inconsistency-based training or scoring?

  3. Can the authors demonstrate that inconsistency separates labeled anomalies from normal nodes beyond degree and relation-density effects?

4 comments

We thank the reviewer for the insightful comments. We respond to the weaknesses and questions below and have revised the manuscript with revision content highlighted. We would appreciate it if you could let us know if our response addresses your concerns.

For clarity, we organize the reviewer's weakness concerns into six points:

  • W1: Lack of a matched multi-relation baseline.
  • W2: Fairness of hyperparameter tuning across baselines
  • W3: Transparency of label Usage and model selection protocol.
  • W4: Set-size effects in cross-relation prediction disagreement.
  • W5: Mechanism of inconsistency-aware training and anomaly separation.
  • W6: Dataset-dependent ablation results and scope of performance claims.

Q1&W2&W3. Model training and hyperparameter selection

R1. We clarify our hyperparameter selection and evaluation protocol here. Following Tang et al. [1], we adopt AUROC and AUPRC as evaluation metrics. Hyperparameters are selected through grid search based on AUROC using ground-truth anomaly labels under a single fixed seed. The selected configuration is then evaluated over five independent runs, with the mean and standard deviation reported. Importantly, anomaly labels are not used for model training or checkpoint selection. RelIn-GAD is trained for a predefined number of epochs without early stopping, and the checkpoint with the lowest unsupervised training loss is selected for evaluation.

Regarding evaluation fairness, each vanilla backbone and its RelIn-GAD-augmented counterpart share identical backbone-specific configurations, with additional tuning restricted to RelIn-GAD-specific hyperparameters. This provides controlled paired comparisons under matched backbone settings. Other baselines retain their original default configurations rather than receiving an equivalent hyperparameter search budget. We have clarified the hyperparameter and checkpoint selection protocols in Appendix K.5 (Line 1337) of the revised manuscript.

[1] Jianheng Tang, Fengrui Hua, Ziqi Gao, Peilin Zhao, Jia Li, GADBench: Revisiting and Benchmarking Supervised Graph Anomaly Detection. NeurIPS 2023.

Q2&W1. Matched multi-relation controlled experiment

R2. To disentangle the benefits of multi-relation reconstruction from those of the proposed adaptive mechanisms, we conduct a controlled comparison involving three configurations:

  • Vanilla: The original single-relation backbone, as used in Table 1 (Line 379).
  • Uniform: Multi-relation reconstruction with uniform relation weights (Wi,r=1RW_{i,r}=\frac{1}{R}), without inconsistency-aware training or inference correction (σi=1,γ=0\sigma_i=1,\gamma=0).
  • Full: The complete RelIn-GAD framework, incorporating both adaptive relation weighting and inconsistency-aware training and scoring.

All multi-relation variants use identical backbone configurations and training settings, with the corresponding RelIn-GAD parameters fixed to those of the main experiments.

Table I: Controlled comparison of multi-relation reconstruction and inconsistency-aware modeling (AUROC / AUPRC, %).

VariantAmazonYelpChiS-FFSDIEEE-CIS
Vanilla83.40 / 23.9361.75 / 23.1963.11 / 26.9967.74 / 16.00
Uniform87.38 / 29.6560.00 / 22.6263.19 / 26.3769.03 / 18.90
Full (ours)88.41 / 36.6662.69 / 24.1564.75 / 29.4969.33 / 18.91

As shown in Table I, Full outperforms Vanilla on all four datasets. More importantly, compared with the matched multi-relation baseline Uniform, Full achieves consistent improvements in both AUROC and AUPRC. Notably, the AUPRC improvements on Amazon and S-FFSD correspond to 23.6% and 11.8% relative gains, respectively. These results indicate that the improvements cannot be attributed solely to multi-relation reconstruction. Instead, adaptive relation weighting and inconsistency-aware mechanisms (Full) provide additional benefits beyond Uniform, with varying magnitudes across datasets. We have added the controlled multi-relation reconstruction comparison in Appendix E.2 (Line 918) of the revised manuscript.

Q3. Inconsistency signals beyond degree and relation-density effects

R3. We thank the reviewer for raising this important question. We clarify that our method does not assume that explicit or implicit inconsistency scores independently distinguish labeled anomalies from normal nodes. Instead, inconsistency is designed to characterize relation-dependent discrepancies, which are incorporated as auxiliary information alongside reconstruction evidence for anomaly detection.

Our preliminary study motivates this design by showing that different relation views produce partially overlapping yet distinct high anomaly-score predictions. Such disagreement does not necessarily indicate anomalous behavior and may also be influenced by relation-specific degree distributions and relation densities. Therefore, our motivation is not that inconsistency independently separates anomalies from normal nodes beyond all structural effects, but that relation-dependent discrepancies provide information that can be modeled within the full detection framework.

The contribution of inconsistency-related components is evaluated through the ablation studies in the manuscript. Accordingly, our claim concerns the utility of inconsistency modeling in conjunction with reconstruction evidence, rather than the standalone discriminative ability or structural independence of individual inconsistency scores.

W4. Size-controlled cross-relation prediction disagreement

R4. We thank the reviewer for highlighting the potential influence of set size on the Jaccard distance analysis. We clarify that our preliminary study primarily examines whether the same detector produces consistent high anomaly-score predictions across different relation views, rather than whether anomalous nodes exhibit greater disagreement than normal nodes. Specifically, we focus on the pairwise overlap of relation-specific Top-KK predictions, as identifying high-scoring candidate anomalies is central to GAD.

To address the reviewer's concern about set-size effects, we additionally conduct a size-controlled analysis using DOMINANT and DONE. For each relation, we select the Top-KK nodes according to their anomaly scores and independently sample KK nodes from the remaining N−KN-K nodes as a random reference. Both sets therefore contain exactly KK nodes, allowing us to examine cross-relation differences under equal set sizes. We set KK to the number of ground-truth anomalous nodes (821 for Amazon and 6,677 for YelpChi) solely for this post-hoc evaluation and compute pairwise Jaccard distances across relation views.

Table II: Size-matched pairwise Jaccard distances of relation-specific Top-KK predictions and Remaining-KK reference sets. Relation indices (1, 2, 3) correspond to U-P-U, U-S-U, and U-V-U for Amazon, and R-U-R, R-T-R, and R-S-R for YelpChi.

ModelDatasetSet1–21–32–3
DOMINANTAmazonTop-KK0.87840.85170.6248
Remaining-KK0.11700.11260.0728
YelpChiTop-KK0.73240.81310.7680
Remaining-KK0.18440.21090.1895
DONEAmazonTop-KK0.80930.80760.5906
Remaining-KK0.09510.10170.0682
YelpChiTop-KK0.91010.89820.8412
Remaining-KK0.24820.24340.2247

As shown in Table II, substantial cross-relation differences persist among Top-KK predictions despite identical set sizes. The pairwise Jaccard distances of Top-KK sets range from 0.5906 to 0.9101 across the evaluated relation pairs, indicating that different relation views identify partially overlapping yet distinct high-scoring candidate anomalies. The Remaining-KK sets provide a size-matched reference rather than a representation of ground-truth normal nodes.

These findings reinforce our original motivation for modeling cross-relation behavioral inconsistency, as different relation views may reveal distinct anomaly-related patterns that provide complementary information for GAD. Importantly, cross-relation disagreement is not treated as a standalone anomaly criterion. Instead, RelIn-GAD models relation-dependent variation alongside reconstruction evidence to capture both shared and relation-specific patterns.

We have clarified the motivation in Sec. 1 (Line 77), revised the preliminary study in Sec. 4.2 (Line 368), and added the size-matched Jaccard analysis in Appendix B (Line 794) of the revised manuscript.

W5. Explanation of the Inconsistency-Aware Training Mechanism.

R5. We clarify the motivation and role of Eq. (6) (Line 312) as follows.

First, cross-relation inconsistency does not necessarily indicate anomalousness, as benign nodes may also exhibit substantial variation due to distinct relation semantics. Thus, Eq. (6) does not explicitly prioritize anomalies but assigns greater optimization emphasis to nodes with complex cross-relation patterns that may be insufficiently captured under uniform weighting. Second, the objective aims to improve relation-aware reconstruction learning, rather than directly enforce anomaly separation. By increasing the contribution of highly inconsistent nodes, Eq. (6) encourages the backbone to better capture difficult relational patterns. This could improve anomaly discrimination when benign relational variations are better reconstructed while genuinely anomalous patterns remain difficult to explain. Third, anomaly detection does not rely solely on the inconsistency-weighted reconstruction objective. The final anomaly score combines attention-weighted reconstruction evidence with a gated explicit structural correction (Eq. (7), Line 321), allowing explicit structural inconsistency to serve as complementary evidence for anomaly detection. We have further clarified the optimization effect and limitations of inconsistency-aware scaling in Appendix F.1 (Line 942) of the revised manuscript.

W6. Effectiveness and scope of applicability

R6. We thank the reviewer for this valuable suggestion. The effectiveness of inconsistency modeling depends on the characteristics of cross-relation disparities in each dataset. On IEEE-CIS, only 0.28% of nodes belong to the implicit-inconsistent category, while 74.91% are relation-consistent, limiting the contribution of implicit inconsistency. Although 24.81% of nodes are explicitly inconsistent, the average explicit score correction is relatively small compared with reconstruction-score variability (0.0211 times its standard deviation). These observations are consistent with the limited changes in aggregate AUROC/AUPRC, including the identical values reported to two decimal places in the original Table 3 (Line 487); they do not imply that node-level scores are identical. We also acknowledge that FreeGAD achieves higher AUPRC on Amazon (69.33% vs. 35.58%). Our primary contribution is not to outperform every specialized method on every metric, but to provide a plug-and-play framework that enhances reconstruction-based graph anomaly detectors by incorporating multi-relation information. As demonstrated in Table 1 (Line 379), RelIn-GAD improves multiple backbones across diverse datasets, although the magnitude of improvement varies. We have expanded the discussion of dataset-dependent ablation results in Sec. 4.5 (Line 451) and added an analysis of the scope and limitations of RelIn-GAD in Appendix H (Line 1080) of the revised manuscript.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer tWnw,

Thank you again for your careful review and valuable suggestions.

As the discussion deadline approaches, we would be grateful to know whether our responses and revisions have addressed your concerns. If you feel they have, we would appreciate it if you could reconsider your rating. We would be happy to clarify any remaining points.

Best wishes,

The Authors

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

This paper proposes RelIn-GAD, a framework for extending reconstruction-based single-relation anomaly detectors to multi-relation graphs. It combines global-to-local relation weighting with structural and reconstruction-based inconsistency signals for node-adaptive training and gated score correction. Experiments on four datasets and seven backbones show improvements in most settings, supporting the practical value of the framework.

Critical strengths
  1. Practical plug-and-play design. The framework adapts seven reconstruction-based GAD backbones to multi-relation graphs while retaining their encoder architectures, demonstrating broad applicability across existing detectors.
  2. Interpretable dual-inconsistency framework. Decomposing anomalies into explicit structural and implicit behavioral divergences is intuitive. The case studies well demonstrate how these complementary signals capture complex anomalies.
  3. Broad empirical validation. The evaluation covers four datasets and 18 baselines. The authors also transparently acknowledge the method's limitations on datasets with consistent relations, strengthening the paper's credibility.
Critical weaknesses
  1. The benefit of the inconsistency taxonomy is not established. Table 3 removes individual components but does not compare the proposed categorization and gating with simpler inconsistency measures.
  2. The asymmetric treatment of explicit and implicit inconsistency lacks empirical justification. Both signals affect training, but only explicit inconsistency contributes directly to the inference score.
  3. The hyperparameter selection protocol is underspecified. RelIn-GAD uses grid search while other baselines use default configurations, without specifying the selection data or criterion.
Questions for authors
  1. With relation weighting held fixed, how does the proposed taxonomy compare with applying degree or reconstruction-error dispersion to all nodes without category-based gating?
  2. Why is implicit inconsistency excluded from the inference score? What happens when it is included alongside explicit inconsistency?
  3. What data and metric are used for hyperparameter selection? Are anomaly labels used, and are selection data separated from evaluation data? How do the results change when competitive baselines receive comparable tuning?
3 comments

We thank the reviewer for the insightful comments. We address the concerns below, with manuscript revisions highlighted.

W1&Q1. Comparison with simple inconsistency measures and category-based gating.

R1. Our taxonomy is not an anomaly classifier. The proposed node taxonomy aims to identify reliable cross-relation evidence for each node rather than directly classify anomalies. Explicit inconsistency captures structural disparities between relations, while implicit inconsistency measures reconstruction disagreement when nodes are represented across all relations. The taxonomy therefore determines which inconsistency signal is used for node-adaptive training.

To evaluate the contribution of category-based selection, we compare five strategies for constructing the training-time inconsistency score. All variants use the same backbone architecture, relation-weighting mechanism, and training configurations, differing only in the inconsistency score used to compute the node-wise scaling coefficients in Eq. (6) (Line 311). Importantly, the inference scoring function remains unchanged (Eq. (7), Line 321) across all variants, isolating the effect of taxonomy-based selection during training.

  • Degree-All: Uses explicit structural inconsistency siExs_i^{\mathrm{Ex}} for all nodes, without category-based selection.

  • Recon-All: Uses implicit reconstruction inconsistency siIms_i^{\mathrm{Im}} for all nodes, without category-based selection.

  • Dual-All: Combines both inconsistency signals for every node, using βtsiEx+(1−βt)siIm\beta_t s_i^{\mathrm{Ex}}+(1-\beta_t)s_i^{\mathrm{Im}}, without category-based gating.

  • Random-Gate: Retains the original category-based scoring formulation but randomly assigns nodes to the explicit-inconsistent, implicit-inconsistent, and relation-consistent categories, preserving the original category sizes.

  • Dual-Gated (ours): Uses the proposed degree-based taxonomy to select the appropriate inconsistency signal for each node, as defined in Eq. (4) (Line 292).

All variants follow the same normalization procedure as Eq. (4), and the resulting inconsistency scores are used exclusively for training-time node weighting.

Table I: Effect of training-time inconsistency selection strategies (AUROC / AUPRC, %).

VariantAmazonYelpChiS-FFSDIEEE-CIS
Degree-All78.84/32.4062.59/24.0463.93/27.3869.00/18.31
Recon-All88.31/35.5562.64/24.1362.94/25.6968.76/18.23
Dual-All76.67/31.5962.62/24.1163.57/26.7468.96/17.72
Random-Gate88.37/36.6062.66 / 24.1663.47/26.4469.08/18.52
Dual-Gated (ours)88.41 / 36.66**62.69 / 24.1564.75 / 29.4969.33 / 18.91

As shown in Table I, Dual-Gated achieves the highest AUROC across all four datasets, demonstrating competitive performance against alternative inconsistency selection strategies. Especially compared with Random-Gate, the improvement is most pronounced on S-FFSD, with relative gains of 2.02% and 11.54% in AUROC and AUPRC, respectively. These observations suggest that the benefit of taxonomy-guided signal selection is dataset-dependent rather than uniformly substantial.

Importantly, this experiment evaluates how inconsistency signals are selected for different nodes, whereas the original component ablations examine the contribution of incorporating inconsistency mechanisms. The two analyses therefore address complementary aspects of our design. We added the taxonomy ablation in Sec. 4.5 (Line 451) of the revised manuscript.

W3&Q3. Hyperparameter selection protocol

R2. We clarify our hyperparameter selection and evaluation protocol here. Following Tang et al. [1], we adopt AUROC and AUPRC as evaluation metrics. Hyperparameters are selected through grid search based on AUROC using ground-truth anomaly labels under a single fixed seed. The selected configuration is then evaluated over five independent runs, with the mean and standard deviation reported. Importantly, anomaly labels are not used for model training or checkpoint selection. RelIn-GAD is trained for a predefined number of epochs without early stopping, and the checkpoint with the lowest unsupervised training loss is selected for evaluation.

Regarding evaluation fairness, each vanilla backbone and its RelIn-GAD-augmented counterpart share identical backbone-specific configurations, with additional tuning restricted to RelIn-GAD-specific hyperparameters. This provides controlled paired comparisons under matched backbone settings. Other baselines retain their original default configurations rather than receiving an equivalent hyperparameter search budget. We have clarified the hyperparameter and checkpoint selection protocols in Appendix K.5 (Line 1337) of the revised manuscript.

[1] Jianheng Tang, Fengrui Hua, Ziqi Gao, Peilin Zhao, Jia Li, GADBench: Revisiting and Benchmarking Supervised Graph Anomaly Detection. NeurIPS 2023.

W2&Q2. Asymmetric treatment of explicit and implicit inconsistency.

R3. Explicit inconsistency is derived from cross-relation structural disparities and provides reconstruction-independent evidence for anomaly scoring. In contrast, implicit inconsistency measures the dispersion of the same relation-wise reconstruction residuals used in the attention-weighted reconstruction score (Eq. (5), Line 303). Although implicit inconsistency captures cross-relation variation, incorporating it directly into the inference score may introduce redundant or less reliable reconstruction evidence. In particular, the attention mechanism may assign low weights to relations with large residuals (e.g., v7769v_{7769}, Line 1062), whereas the implicit dispersion score treats these residuals equally.

During training, implicit inconsistency serves a different purpose: it modulates node-wise optimization through a stop-gradient scaling coefficient (Eq. (6), Line 312). To directly examine the effect of including implicit inconsistency alongside explicit inconsistency at inference, we compare two scoring strategies using the same trained full RelIn-GAD model.

  • Explicit+Implicit: Extends Eq. (7) (Line 321) by adding a gated implicit correction, si=ℓcombined,i+γsiExmEx,i+γsiImmIm,is_i=\ell_{\mathrm{combined},i}+\gamma s_i^{\mathrm{Ex}}m_{\mathrm{Ex},i}+\gamma s_i^{\mathrm{Im}}m_{\mathrm{Im},i}.

  • Explicit-only (ours): The original inference score defined in Eq. (7), si=ℓcombined,i+γsiExmEx,is_i=\ell_{\mathrm{combined},i}+\gamma s_i^{\mathrm{Ex}}m_{\mathrm{Ex},i}.

Here, mEx,im_{\mathrm{Ex},i} and mIm,im_{\mathrm{Im},i} are the taxonomy-based indicators defined in Sec. 3.2.2 (Line 240). We use the same correction coefficient γ\gamma for both inconsistency signals, with its value fixed to that of the original model, ensuring that the comparison isolates the effect of adding implicit inconsistency at inference.

Table II: Effect of including implicit inconsistency in inference scoring (AUROC / AUPRC, %).

VariantAmazonYelpChiS-FFSDIEEE-CIS
Explicit+Implicit87.20/32.6161.95/23.5063.73/26.9669.16/18.48
Explicit-only (ours)88.41 / 36.6662.69 / 24.1564.75 / 29.4969.33 / 18.91

Across all four datasets, Explicit-only outperforms Explicit+Implicit in both AUROC and AUPRC, supporting the decision to omit implicit inconsistency from the inference correction under this fixed-γ\gamma comparison. We have clarified the asymmetric treatment of inconsistency signals in Sec. 3.3 (Line 323) and added the inference scoring comparison in Appendix F.2 (Line 990) of the revised manuscript.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer dpqp,

Thank you again for the time and effort devoted to reviewing our paper.

As the discussion deadline approaches, we would appreciate it if you could let us know whether our responses and revisions have addressed your concerns. We would be happy to clarify any other points.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Rebuttal

Our paper proposes RelIn-GAD, a plug-and-play framework for unsupervised multi-relation graph anomaly detection. By distinguishing explicit and implicit cross-relation inconsistencies, RelIn-GAD integrates adaptive relation weighting, inconsistency-aware training, and inference correction to enhance existing reconstruction-based detectors. Extensive experiments demonstrate its effectiveness across diverse datasets and backbones.

We sincerely thank the reviewers for their insightful feedback, which has significantly strengthened our paper and improved its readability. We have carefully addressed the raised concerns and summarize the key revisions below:

  • Validation of methodological designs (Reviewers dpqp, KY2a): We have added ablations comparing taxonomy-guided selection with simpler inconsistency measures and random gating. We also evaluated alternative inference scoring strategies, providing further empirical support for our taxonomy and asymmetric treatment of inconsistency signals.

  • Strengthened motivation and mechanism analysis (Reviewer tWnw): We have conducted size-matched Jaccard distance analyses to verify cross-relation prediction disagreement. We further clarified the auxiliary role of inconsistency signals and explained how inconsistency-aware training facilitates the modeling of challenging relational patterns.

  • Clarification of experimental protocols (Reviewers dpqp, tWnw): We have detailed the hyperparameter selection, label usage, checkpoint selection, and evaluation procedures, including the use of anomaly labels for hyperparameter tuning and differences in tuning budgets, improving experimental transparency.

  • Performance analysis and technical clarification (Reviewers tWnw, KY2a): We have explained dataset-dependent ablation results, AUROC values below 50%, and technical differences from UMGAD and SFGA. We also acknowledged cases where competing methods outperform ours and clarified the applicability of RelIn-GAD.

  • Additional baselines and controlled experiments (Reviewers tWnw, KY2a): We have introduced a matched multi-relation baseline to isolate the contributions of our proposed mechanisms and added comparisons with heterogeneous graph methods MUG and MVMAE, further demonstrating the effectiveness of RelIn-GAD.

Overall, we believe our comprehensive responses and manuscript revisions have substantially strengthened the paper and addressed the reviewers' major concerns. Although the discussion phase concluded before we could receive further feedback, we respectfully request that the Area Chair give full consideration to the additional evidence, clarifications, and manuscript improvements presented in our rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.