acceptodds
Under review as a conference paper at ICLR 2027

Granularity-aware hierarchical information alignment for multi-granularity Graph Anomaly Detection

Abstract

Multi-granularity graph anomaly detection is a recently emerging issue that aims to design a unified solution for the detection of multi-granularity anomalies in graphs. Existing methods usually employ subgraph sampling and unified training objectives, achieving promising performance. However, they ignore hierarchical information in latent representations, making the latent-space evolution less discriminative. Moreover, they fail to explore interdependencies across granularities in specific detection tasks, which could further regularize the representations. To address these limitations, we propose a granularity-aware hierarchical information alignment framework for multi-granularity GAD (GaHIGAD). Specifically, GaHIGAD explores latent evolutionary information about hidden representations based on Information Bottleneck (IB) theory. In addition, GaHIGAD integrates cross-granularity representations to exploit latent consistency dependencies and achieve representation regularization. Extensive experiments on comprehensive public datasets demonstrate the effectiveness and advancement of GaHIGAD compared to state-of-the-art baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-granularity graph anomaly detection, unsupervised learning, information bottleneck, contrastive 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_oNCkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses multi-granularity graph anomaly detection with a single unified model. The proposed GaHIGAD applies an Information Bottleneck loss to every intermediate encoder layer, and adds a consistency loss aligning representations both across granularities and between fused and per-granularity representations, estimated via InfoNCE and MINE. Experiments on 14 datasets, including zero-shot transfer and robustness tests, report consistent gains over recent unified GAD baselines.

Critical strengths
  1. The empirical validation is relatively comprehensive: it covers 14 public datasets across both single-graph and multi-graph settings, including cross-granularity zero-shot transfer, noise robustness, and component-level ablations, showing consistent advantages over recent baselines.
  2. The granularity-aware gated fusion and the explicit separation of per-granularity encoders are simple and interpretable.
Critical weaknesses
  1. The paper's main novelty is applying the IB loss to all intermediate layers, but no ablation compares this with applying the IB loss to the final layer only. In addition, with β=0.001 this term becomes too small; although experiments were conducted, the analysis is lacking.
  2. Section 4.3 claims that three different strategies are used to estimate the mutual information, but only two are actually described.
  3. No code is released.
Questions for authors
  1. Could you describe the training procedure in detail, which terms are maximized and which are minimized? 2.. Could you provide a more detailed analysis of the estimators, in particular a lower-bound/upper-bound analysis for each mutual information term?
  2. In edge-level and graph-level tasks, how are the samples indexed and aligned? Such as, y is indexed by edges/graphs while h is indexed by nodes, which pairs (y,h ) enter the mutual information terms?
  3. Will you release the code?
Reviewer_fnWkOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new granularity-aware hierarchical information alignment framework for multi-granularity graph anomaly detection. By incorporating an information bottleneck mechanism, the framework explores complex hierarchical information within graph representations; it further models dependencies across different granularities through cross-granularity representation fusion and consistency adjustment. The authors evaluate the effectiveness of the proposed method through extensive comparative and ablation experiments.

Critical strengths
  1. The paper is well organized, and the proposed method is clearly described and easy to follow.
  2. The authors design a consistency adjustment strategy and combine it with the information bottleneck loss to consider both inter-granularity latent consistency and cross-granularity correlations in granularity-specific detection tasks.
  3. The authors conduct extensive experiments on multiple datasets to demonstrate the effectiveness of the proposed method.
Critical weaknesses
  1. The Information Bottleneck (IB) theory has already been extensively researched and applied across various graph learning techniques, including graph-level anomaly detection; therefore, simply introducing it into a multi-granularity anomaly detection task for representation extraction offers limited novelty.
  2. Although the authors design separate encoders for different granularities, the joint optimization objective incorporates cross-granularity consistency constraints, which may cause the encoders for different granularities to influence one another. Consequently, I am skeptical as to whether these encoders can effectively retain their ability to capture granularity-specific anomaly patterns.
  3. The overall framework design is relatively straightforward, relying primarily on GNNs and the IB theory for representation learning, followed by the construction of multi-granularity representations through node representation differencing and pooling operations. There is a lack of deeper consideration or exploration regarding how to balance shared information with the specific requirements inherent to multi-granularity tasks.
Questions for authors
  1. Although the authors designed separate encoders for different granularities, the losses associated with these granularities may influence one another during joint optimization. How do the authors prevent interference between tasks at different granularities and ensure that each encoder effectively captures anomaly patterns specific to its respective granularity?
  2. How do the authors evaluate the information complementarity between representations of different granularities? It remains unclear whether the performance gains truly stem from complementary information across granularities, or if they simply arise from fusing representations from multiple encoders—an effect akin to model ensembling—potentially improving both model stability and detection performance.
Reviewer_CDuCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes GaHIGAD, a single framework for node-, edge-, and graph-level anomaly detection that combines gated fusion, a multi-layer information bottleneck, and a consistency loss. It reports large gains on 14 datasets. However, the objective's signs, how anomaly scores are computed, and the value of multi-layer IB need clarification.

Critical strengths
  1. The problem is useful in practice. The paper detects anomalies at the node, edge, and graph levels in one framework, and uses the links between these levels to learn better representations. This is a well-motivated setting compared with training a separate detector for each level.
  2. The design is clear. The branches for different levels share information through gated fusion. An information bottleneck (IB) loss is applied to several hidden layers, and a consistency loss ties each branch to the fused representation (Eqs. 2–9). The main technical contribution is the layer-wise constraint and the way information is combined across levels. The GNN, the gated fusion, and the MI estimator are standard parts.
  3. The experiments are broad. The paper uses 14 datasets in both single-graph and multi-graph settings, and reports the mean and standard deviation over five seeds. It also includes cross-level transfer, ablations on the loss and the fusion method, sensitivity analysis, and noise robustness. The gains in Tables 1–2 are large. For example, node-level AUROC on Amazon is 94.8, while the best baseline gets 87.2. I think these results should be checked again once the method details are clarified.
Critical weaknesses
  1. The signs and the optimization direction of the main objective are unclear. Eq. 11 writes the negative of the standard Donsker–Varadhan lower bound used in MINE [1] and calls it mutual information. Eqs. 8–9, however, add the consistency terms with a positive sign. If Eq. 11 is read as an estimate of negative MI and Eq. 9 is maximized, then the three goals (compressing the input, keeping class information, and improving consistency) do fit together. The paper should say which reading is intended. It should also say in which direction the full objective and the MI critic are each optimized. Separately, the sum in the denominator of Eq. 10 does not depend on the index, so as written the expression is a constant. This should be fixed.
  2. It is not clear how the class outputs become anomaly scores. Eq. 6 defines y_i as a class probability vector, and the consistency loss is meant to provide a training signal without labels. However, the paper does not explain three things: how the two outputs come to mean "normal" and "anomalous", how trivial solutions where both outputs always agree are avoided, and how the final anomaly score is computed. Algorithm 1 returns Scores but does not say how they are obtained.
  3. The benefit of applying IB to multiple layers is not tested directly. Tables 3–4 replace IB with reconstruction or contrastive losses. This shows that the IB loss helps overall, but it does not show how much comes from adding the constraint to the middle layers. A comparison that keeps the encoder fixed and only changes the number of layers with IB would answer this more directly.
  4. Some implementation and baseline details are missing. The paper gives the training ratios, short baseline descriptions, how edge-level detection is done by concatenation, and how edges crossing the train/test split are removed. It does not say: how the MI critic is trained; how baselines that originally use supervision were changed; how edge anomaly probabilities are turned into labels for evaluation. Also, the graph-level description in Appendix C ends mid-sentence, and the equation numbers in Algorithm 1 do not match the main text.

[1] Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajeswar, Sherjil Ozair, Yoshua Bengio, Aaron Courville, and R Devon Hjelm. Mutual Information Neural Estimation. In Proceedings of the 35th International Conference on Machine Learning (ICML), PMLR 80, 2018.

Questions for authors
  1. Please give the full objective in Eqs. 7–11 as used in the code. Is it minimized or maximized, is the MI critic trained in alternation, and what are the positive/negative samples and the correct denominator in Eq. 10?
  2. How do the two outputs in Eq. 6 get their "normal" and "anomalous" meaning, how are trivial solutions where both always agree avoided, and how are the final Scores computed?
  3. With the encoder and training budget fixed, how does performance change if IB is applied only to the last layer?
  4. How were baselines that originally use supervision, such as SGAD-DCR, adapted? Appendix C gives p(e_ij) = Avg(p_i, p_j): where does p_i come from, and how are edge probabilities turned into evaluation labels?
1 comment

I thank the authors for their clarifications and additional experiments. These revisions address my main concerns, and I have raised my score. For the final version, please check for a possible typo in Eq. (10), where the same representation appears in both arguments of the mutual-information term, and briefly clarify how the conditional model used in CLUB is obtained or fitted.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.