- 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
-
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.
-
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
- 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
-
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?
-
How are the dynamic aggregation weights in Eq. 8 updated?
-
How does Eq. (9) ensure that granularity-specific prototypes are sufficiently different instead of highly overlapping?
-
Are the node-, edge-, and graph-level results obtained from one jointly trained model or from separately trained models?