acceptodds
Under review as a conference paper at ICLR 2027

Quantization-conditioned Normalizing Flows for Multi-granularity Graph Anomaly Detection

Abstract

Multi-granularity Graph Anomaly Detection aims to design a unified framework for detecting abnormal patterns at node, edge, and subgraph granularities, occupying an important position in academia and industry. Existing methods usually focus on exploiting a unified representation space for different granularities, while overlooking the distinct characteristics and anomaly patterns across granularities, leading to poor identification ability. To address these limitations, a quantization-conditioned normalizing flow framework for multi-granularity GAD (QNMGAD) is proposed to tackle this challenge. On the one hand, QNMGAD introduces prototype quantization to capture discriminative semantics across granularities and utilizes residual patterns quantization to extract latent semantics within a specific granularity. On the other hand, QNMGAD introduces conditioned normalizing flows based on granularity prototypes to model the distribution of normal patterns and calculate density estimation to identify anomalies. Extensive experiments compared with state-of-the-art baselines show the superiority of QNMGAD.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
multi-granularity graph anomaly detection, unsupervised learning, normalizing flow

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_m5zYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The authors address granularity inconsistency in graph anomaly detection and propose QNMGAD, which surprisedly integrates prototype quantization with normalizing flows to capture discriminative representations, instead of using traditional reconstruction- or contrastive-based paradigms. Some technical details are very enlightening, such as granularity-specific residual pattern quantization and prototype-conditioned normalizing flows. The architecture figure is clean and easy to follow. The experiments show the superiority of QNMGAD over baselines.

Critical strengths
  1. The paper points out a real weakness in existing unified-representation multi-granularity GAD methods: they tend to blur differences across granularities. Modeling granularity-specific normal patterns is a fresh idea, and the motivation is clear. It is also nice that the authors do not just force all granularities into one shared space.

  2. The QNMGAD framework is well put together. The three modules connect logically, and the math is complete. The systematic ablations are a strong plus, since they clearly show what each component brings to the table.

Critical weaknesses
  1. The connection between prototype quantization and granularity discrimination can be more explicit. Eq. (9) mainly minimizes the distance between the aggregated prototype and the representation of each granularity. However, there is no explicit objective to separate prototypes selected by different granularities. Therefore, it is unclear why this formulation can necessarily avoid semantic overlap across granularities.
Questions for authors
  1. The paper argues that existing unified multi-granularity methods overlook granularity-specific anomaly patterns. However, the examples in the Introduction mainly show that node and edge anomalies can have different semantics, rather than explaining why existing methods cannot capture such differences. Can author make further illustration of this issue?

  2. How are the dynamic aggregation weights in Eq. 8 updated?

  3. How does Eq. (9) ensure that granularity-specific prototypes are sufficiently different instead of highly overlapping?

  4. Are the node-, edge-, and graph-level results obtained from one jointly trained model or from separately trained models?

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

The paper looks at multi-granularity graph anomaly detection. This problem seems to be interesting because it asks for anomalies at node, edge, and graph/subgraph levels at the same time, which is not easy. The authors make a fair point: many existing unified frameworks put different granularities into one shared space, so granularity-specific semantics can get lost. Authors propose QNMGAD based on quantization-conditioned normalizing flows. The design has two parts. One captures cross-granularity differences and fine-grained semantics with prototype quantization and residual pattern quantization. The other uses prototype-conditioned normalizing flows to model normal data and estimate anomalies by density. They test on 14 single-graph and multi-graph datasets, and QNMGAD beats strong baselines pretty consistently. The experimental validation is comprehensive and convincing.

Critical strengths

1.The target problem is quite advanced and have great significance, and the framework has certain contributions.

2.The experimental section is exceptionally thorough, evaluating the method across 14 diverse single-graph and multi-graph datasets. It includes comprehensive comparisons with numerous state-of-the-art baselines, rigorous cross-task transferability tests, detailed ablation studies, sensitivity analyses, and robustness evaluations against noise, which strongly substantiate the claims made in the paper.

Critical weaknesses
  1. This article designs a total of three loss functions, however, the author did not clarify what kind of interactions exist between them.

  2. In the Transferability experiment, the author's experimental details were insufficient.

Questions for authors
  1. Can the author further clarify what kind of relationship exists between the loss functions?
  2. Can the author further clarify the details of the transfer experiment, as the sample size of training granularity may directly affect transfer performance. Additionally, will using different proportions of training data have an impact on the transfer effect?
Reviewer_71V3OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work studies multi-granularity GAD, which is an recently emerging topic. Authors reveal that existing methods can suffer from semantic overlap and information loss when they force everything into one rigid shared space. They propose QNMGAD, which takes a different angle for this issue. Instead of fully unifying the space, it models granularity-specific normal patterns into granularity-specific spaces through discrete quantization and conditional generation. It also designs a hierarchical quantization mechanism with prototype quantization and prototype-conditioned normalizing flows for robust anomaly scoring. Therefore, the technical contributions are clear and solid. The experimental results also support cross-task transferability.

Critical strengths

1.Unlike prior multi-granularity GAD methods that naively enforce a completely unified local space, this paper insightfully identifies the granularity inconsistency problem, revealing that anomaly patterns across different granularities can exhibit conflicting semantic characteristics. Recognizing and explicitly modeling these discrepancies represents a conceptually refreshing and important shift for unified graph anomaly detection. 2.The combination of multi-granularity prototype quantization (for global granularity discrimination) and granularity-specific residual pattern quantization (for fine-grained intra-granularity semantics) is highly creative. This design successfully bridges coarse-grained concept discrimination with fine-grained feature refinement without relying heavily on high-quality labels.

Critical weaknesses

1.The paper need to further clarify the challenges introduced by granularity inconsistency. If anomaly patterns are inconsistent across granularities, it would be useful to clarify what kinds of inconsistency are considered. The current paper mainly emphasizes granularity-specific modeling but provides limited discussion on this trade-off. 2. The multi-granularity training setting should be clarified. From Eqs. (19) and (20), it is not fully clear whether one model jointly detects anomalies at different granularities or whether separate models are trained for different tasks.

Questions for authors

1.Along with Weakness 1, the author has defined a new direction for multi granularity anomaly detection. However, can the author further explain its practical significance from a practical application perspective. 2. Could you further illustrate how representations of different granularities affect anomaly scores and whether the contribution of each granularity can be reflected by the activation degree of the corresponding prototype set? 3. Please see Weakness 2. 4. In Table 5, authors should clarify which metric is reported.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.