- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies unsupervised cross-domain, cross-granularity graph anomaly detection and proposes CgGFAD, which combines node-level representation alignment, Wasserstein-based granularity discrimination, cluster-based self-distillation, and prototype generalization to detect anomalies without retraining on target domains. Experiments on 14 datasets across multiple transfer directions report performance improvements over existing baselines.
Critical strengths
- This work provides clear motivation, application background, and a framework.
- The cross-granularity generalization proposed in this work is sound.
- Experiments cover multiple datasets, and the results seem good.
Critical weaknesses
- The method uses source-domain representations at both the training and test granularities, but does not clarify how they are adapted to cross-granularity tasks.
- The practical necessity of optimizing the anomaly detection objective at only one granularity is insufficiently justified.
- The full-model results in Table 8 duplicate those in Table 6, although the tables indicate different training and test dataset assignments.
- The connection between granularity separation and knowledge transfer is insufficiently justified.
Questions for authors
- Do all baselines have access to the same source-domain granularity information as CgGFAD, and how are they adapted to cross-granularity tasks?
- When representations at both granularities are available in the source domain, why is the anomaly detection objective optimized at only one granularity, and what substantive innovation does this setting offer beyond GFM-UAD?
- Why do the full-model results in Table 8 match those in Table 6 despite different training/test dataset assignments, while differing from Table 2 under the same setting?
- Please further explain how granularity separation facilitates cross-granularity transfer of anomaly-related knowledge, and justify the use of means and standard deviations to characterize representations and construct the MMD loss.
- Node features may differ in dimensionality and semantics across datasets; how does CgGFAD handle these differences and achieve cross-dataset feature alignment?