- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The authors investigate test-time graph anomaly detection under normality shift, where evolving benign patterns introduce structural contamination and prediction bias from the source detector. The proposed method PUPPY combines perturbation-aware structure denoising with adaptive proxy refinement to tackle both problems. Evaluation is conducted under controlled and natural temporal shift protocols across diverse benchmarks.
Critical strengths
This work treats test-time GAD under normality shift as a joint problem of structural contamination and prediction bias, which is a sound and well-motivated perspective. It addresses the two problems through perturbation-aware structure denoising and adaptive proxy refinement, yielding a coherent framework.
Critical weaknesses
- Teacher initialization selects anchor normals by the smallest source anomaly margins (Eq. 5), implicitly assuming the frozen source detector remains discriminative on the target graph. Under severe normality shift, evolved normals may carry inflated source margins, making the anchor set noisy from the outset. The paper provides no analysis of anchor purity or how proxy quality degrades as shift intensity increases.
- (Eq. 13) regularizes the student toward source predictions throughout adaptation. Under normality shift, source predictions on evolved normals introduce the very bias the method aims to correct. If the source reference is misaligned, this term may reinforce rather than correct student errors, propagating through EMA into both proxy targets and structure denoising. The paper does not analyze this interaction.
Questions for authors
1.When the source model gives a high abnormal score for the absence of normal patterns, how can the method ensures that the BottomK anchor points still represent the true normality of the target and prevent the teacher agent from adapting from the wrong anchor points?
2.When the source detector is overall unreliable under strong drift, the source reference can not provide effective correction, and how should source reference failure be detected and dealt with?
3.Relying solely on AUROC and AUPRC may not be sufficient to comprehensively evaluate the effectiveness of PUPPY.It is noted that the baseline model and dataset selected in this article have almost no differences from those used in Panetal.,2026. It is necessary to supplement more current relevant work for baseline comparison.
4.The article is missing the system optimization process for numerous hyperparameters of PUPPY. How does the parameter tuning cost of the method affect its suitability for actual deployment? Please supplement the relevant experiments on the impact of parameter adjustment on the performance of the method.
2 comments
Response to Reviewer 1Csp (Part 2/2)
Q5: Extending the evaluation with additional metrics and recent baselines.
A5: Thank you for this suggestion. We have added recent graph anomaly detectors RHO [1] and GraphNC [2], alongside graph TTA methods Matcha [3] and TSA [4]. We also report Recall@k to complement AUROC and AUPRC by measuring anomaly retrieval among the top- nodes, where is the number of anomalies in the test split.
The table reports means (%) over six datasets, additionally averaged over three backbones for the TTA methods.
| Method | Mean AUROC | Mean AUPRC | Mean Recall@k |
|---|---|---|---|
| RHO | 66.40 | 27.84 | 29.83 |
| GraphNC | 59.32 | 17.50 | 18.45 |
| TSA | 64.26 | 25.37 | 23.66 |
| Matcha | 67.36 | 25.04 | 25.21 |
| PUPPY | 73.81 | 36.29 | 37.05 |
PUPPY maintains its lead over these recent baselines across all three metrics. Its higher Recall@k further shows that it retrieves more true anomalies among the top-ranked nodes, helping prioritize them for inspection.
Following prior work [5], we extend our evaluation to MGTAB [6], a Twitter account detection benchmark with human/bot labels. Each cell reports source detector → PUPPY (%).
| Backbone | AUROC | AUPRC | Recall@k |
|---|---|---|---|
| GCN | 73.96 → 75.69 | 60.34 → 62.65 | 52.53 → 54.12 |
| BWGNN | 85.76 → 87.63 | 71.07 → 73.09 | 65.22 → 68.28 |
| GHRN | 86.44 → 87.82 | 71.79 → 73.06 | 66.73 → 67.60 |
On MGTAB, PUPPY also achieves the best performance among the evaluated methods, supporting its broad applicability. We will include the full comparisons in the revised manuscript.
Q6: Hyperparameter sensitivity and tuning effort.
A6: Thank you for raising this practical question. We extend the sensitivity analysis to five additional hyperparameters across four tasks. 85.53% of tested parameter changes keep AUROC within 1.00 percentage point of the reference, showing stable performance across most evaluated settings.
| Parameter | Settings within 1 AUROC point (%) |
|---|---|
| Edge-allocation temperature | 100.00 |
| Drift clipping threshold | 81.25 |
| EMA momentum | 93.75 |
| Teacher update interval | 93.75 |
| Proxy supervision ratio | 50.00 |
For example, on BWGNN–T-Social, varying the teacher update interval from 10 to 30 epochs yields 84.58%–84.80% AUROC, while proxy ratios from 0.60 to 1.00 yield 84.65%–84.75%.
These stable ranges reduce the need for fine-grained tuning of most parameters. In a further search on GHRN–MGTAB with the same number of additional candidate configurations per method, PUPPY also achieves the highest AUROC among the compared TTA methods. These results support PUPPY’s practical applicability by demonstrating stable performance across a range of parameter settings. We will provide a more detailed discussion in the revised manuscript.
Thank you again for your positive assessment and thoughtful suggestions. We hope these responses and additional analyses address your concerns. We welcome further questions and would be happy to provide any additional clarification.
[1] Guoguo Ai et al. Semi-supervised graph anomaly detection via robust homophily learning. Advances in Neural Information Processing Systems, 38:133988–134013, 2026.
[2] Guolei Zeng et al. Normality calibration in semi-supervised graph anomaly detection. arXiv preprint arXiv:2510.02014, 2025.
[3] Wenxuan Bao et al. Matcha: Mitigating graph structure shifts with test-time adaptation. In International Conference on Learning Representations, 56906–56944, 2025.
[4] Hans Hao-Hsun Hsu et al. Structural alignment improves graph test-time adaptation. arXiv preprint arXiv:2502.18334, 2025.
[5] Buyun He et al. Boosting bot detection via heterophily-aware representation learning and prototype-guided cluster discovery. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining Vol. 2, 860–871, 2025.
[6] Shuhao Shi et al. MGTAB: A multi-relational graph-based Twitter account detection benchmark. Neurocomputing, 647:130490, 2025.
Response to Reviewer 1Csp (Part 1/2)
A1 & A3: Thank you for raising this concern. BottomK selects the nodes the source detector considers most likely to be normal. Even when some evolved normals receive inflated anomaly scores, the selected nodes can still provide a mostly normal starting set for the teacher. We test this directly by measuring the true anomaly rate among the selected anchors.
On the standard target graphs, anchor anomaly rates are below the test-set baselines in 23 of 24 tasks, with an average reduction of 5.27 percentage points over all tasks. 18 tasks achieve at least 97.00% normal anchor purity.
We also vary evolved-normal retention from 0% to 100% across all 24 tasks, keeping each task’s anchor ratio ρa fixed. Representative anchor anomaly rates (%) are shown below.
Across all five levels, 23 of 24 tasks maintain anchor anomaly rates below their test-set baselines, and 17 tasks vary by at most 1.00 percentage point. The selected anchors therefore remain a useful normal reference across these shifts.
The anchors provide a normal starting point for the teacher, while proxy supervision also reaches nodes outside this set. Drift refinement updates their targets as the student adapts, improving proxy quality and enabling revised supervision for initially excluded evolved normals (see our response to reviewer GFqY, Q1 & Q5, https://openreview.net/forum?id=ZXvQ3FOvhr¬eId=vB1U5Jyw6U).
A2 & A4: We appreciate this important concern. Source regularization can pull evolved normals toward incorrect source scores, but the joint objective in Eq. (14) also learns from refined proxy targets that can correct those judgments. Differentiating Eq. (13) gives
{\partial m_{i,\mathrm{orig}}^S} =\frac{1}{N} \left[\sigma\left(m_{i,\mathrm{orig}}^S\right)-p_i^s\right].$$ This gradient pulls the original-graph prediction toward the source probability $p_i^s$. When a selected evolved normal receives a refined proxy target below its current view prediction, proxy supervision in Eq. (11) pushes that prediction downward. Consistency in Eq. (12) links the two views, and EMA incorporates the resulting student updates into the teacher. The teacher and its denoising estimates can therefore evolve with target adaptation. We test whether retaining the source reference helps correct evolved-normal errors by removing source regularization from both teacher initialization and student adaptation. Across the 17 tasks where it is active, retaining it improves mean AUROC and AUPRC by 7.10 percentage points each. The examples below also show that retaining the source reference helps PUPPY correct errors on evolved normals. We report their false-positive rates at 80% anomaly recall (%, lower is better). | Task | Source | PUPPY w/o source regularization | Full PUPPY | |---|---:|---:|---:| | GCN–Questions | 95.73 | 85.85 | 68.80 | | BWGNN–Questions | 92.81 | 81.68 | 49.11 | | GHRN–YelpChi | 90.04 | 79.39 | 67.76 | | GHRN–Tolokers | 88.25 | 82.13 | 78.09 | Full PUPPY substantially reduces false alarms while retaining source regularization. Removing the constraint increases them in these examples, showing that a source reference can support adaptation even when its predictions on evolved normals require correction. Although source regularization generally improves performance in our experiments, our implementation allows this soft constraint to be weakened or disabled in exceptional cases where it persistently harms adaptation.