- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes BNFlow, a guided flow matching framework for graph-level anomaly detection that generates informative pseudo-anomalies using only normal training graphs. The method combines flow matching toward normal graph representations with discriminator-guided refinement to produce challenging synthetic negatives. A relative velocity budget constrains refinement, while ranking-based feedback adaptively adjusts guidance strength throughout training. The paper also provides theoretical analyses that characterise the local effect of discriminator guidance and bound deviations from normal latent references. Experiments on eight graph benchmarks show competitive performance against existing graph anomaly detection methods, supported by ablation studies, parameter-sensitivity analyses, and computational-efficiency comparisons.
Critical strengths
-The paper addresses an important challenge in graph-level anomaly detection: generating synthetic negatives that remain informative as the detector improves. Combining normal-reference transport with discriminator-guided refinement offers a coherent way to balance resemblance to normal graphs with detection difficulty.
-The paper provides mathematical analyses showing that discriminator-guided refinement maximises the first-order increase in synthetic-negative classification loss under a relative velocity constraint. It also establishes a conditional bound on deviations from normal latent references, providing useful theoretical motivation for the proposed sampling strategy.
-BNFlow achieves the highest reported mean AUROC across all eight benchmarks and the highest mean F1-score on seven datasets. The ablation studies demonstrate substantial contributions from discriminator guidance and flow matching, while the reported computational comparisons indicate considerable training-time improvements over the generative baseline AGDiff.
Critical weaknesses
-The framework combines established techniques, including flow matching, discriminator-gradient guidance, constrained adversarial refinement, and adaptive feedback. Although the integration is reasonable, the paper does not sufficiently clarify how the proposed guidance mechanism fundamentally differs from existing classifier-guided generation or adversarial negative synthesis. The paper needs stronger comparisons with alternative guidance strategies to substantiate its methodological contribution.
-Proposition 1 characterises a local first-order optimisation property, while Proposition 2 bounds deviation from a normal latent reference in terms of velocity errors and accumulated guidance. However, these results do not guarantee that generated samples are sufficiently anomalous, semantically meaningful, or consistently informative for downstream detection. The theoretical connection between the proposed guidance and improved anomaly detection performance therefore remains limited.
-The method aims to generate informative synthetic negatives that capture subtle deviations from normal patterns, but the analysis primarily examines latent-space proximity and classification performance. Figure 6 indicates that approximately 98% of generated samples become normal-like at longer sampling horizons, with uneven coverage across normal prototypes. This raises concerns about insufficient diversity, potential concentration around specific normal modes, and whether the generated samples adequately represent different anomaly types.
-Although the reported results are generally strong, the improvements over AGDiff are relatively small on some datasets, and the paper does not provide paired statistical significance tests to establish their reliability. Additionally, the reported 230–800× training-time advantage over AGDiff is substantial but requires a clearer breakdown of preprocessing, autoencoder pretraining, generation, and joint optimisation costs to establish whether the comparison reflects equivalent computational workloads.
Questions for authors
-What are the key technical differences between the proposed bounded discriminator guidance and existing classifier-guided generation or constrained adversarial negative synthesis methods? Can the authors provide controlled comparisons using alternative guidance mechanisms within the same flow-matching framework to isolate the proposed strategy's contribution?
-How does the method ensure that generated samples remain informative pseudo-anomalies rather than becoming nearly indistinguishable from normal graphs? Given the uneven prototype coverage observed in Figure 6, can the authors provide quantitative analyses of synthetic-negative diversity, coverage across normal modes, and detection performance for different anomaly types?
-Can the authors provide paired statistical significance tests against AGDiff and clarify whether all methods use identical data splits and evaluation protocols? Additionally, does the reported 230–800× training-time improvement include autoencoder pretraining, flow pretraining, and joint optimisation for both methods, and how does performance compare under matched computational budgets?
4 comments
Response to Reviewer iCKC (2/4)
In the following, we clarify weakness 2 concerning the scope of the theoretical results and their connection to detection utility.
2. Scope of the Theoretical Results & Connection to Detection Utility. (W2)
Propositions 1–2 establish locally optimal negative-loss refinement under node-wise budgets and bound normal-reference deviation, without directly guaranteeing semantic anomalousness.
Maintaining Distinguishable Pseudo-anomalies: Although guidance locally increases the normality logit , four mechanisms are designed to reduce the risk of generated samples becoming indistinguishable from normal graphs: (i) Eq. (4) constrains each node's guidance velocity by , limiting the magnitude of the local normality-increasing refinement. (ii) In Eq. (9), the synthetic-negative loss updates both the flow field and detector. Its derivative encourages lower synthetic normality logits, counterbalancing sampling-time refinement toward higher logits. (iii) Eqs. (7)–(8) increase when negatives become too easy to distinguish and reduce it toward its allowed lower bound when the smoothed normal–synthetic separability score falls below . (iv) Where is configured, Eq. (6) stops sampling at an intermediate transport state, providing an additional means to control resemblance to normal references.
Connection to Detection Utility: Appendix B.4 shows that identical normal and generated distributions yield an optimal BCE prediction of under equal class weights, motivating distinguishable negatives. Item 3 shows that guided pseudo graphs remain distinguishable but are harder to separate than unguided samples under the evaluated scores. Items 1.2 and 3 further support detection gains through controlled guidance and anomaly-type comparisons.
Response to Reviewer iCKC (3/4)
In the following, we address weakness 3 and question 2 regarding pseudo-anomaly distinguishability, diversity, and anomaly-type coverage.
3. Generated Pseudo-anomaly Distinguishability, Diversity & Anomaly-type Coverage. (W3, Q2)
Thank you for the insightful question. In Figure 6, "normal-like" denotes proximity under the specified distance threshold in the mean-pooled frozen VGAE space. As reported in Appendix C.2, 71 of the 73 real test anomalies also fall within this region. Thus, the approximately 98% normal-like proportion of generated samples does not mean that they are indistinguishable from normal graphs. To assess their distinguishability and diversity more directly, we added the following analyses on COX2.
Generated Pseudo-anomaly Distinguishability: On COX2, we compared reconstructed normal test graphs with pseudo graphs generated by trained BNFlow models with and without guidance. All graphs used the same frozen VGAE decoder and discretization. For each feature set—atom composition, edge endpoint atom-type pairs, WL patterns, and Laplacian spectra—we scored graphs by their mean distance to five nearest reconstructed normal training graphs, excluding self-matches. AUC compares each group against reconstructed training normals, with 0.5 indicating chance-level discrimination and 1.0 perfect separation.
| Feature | Normal test reference | BNFlow pseudo graphs | Unguided pseudo graphs |
|---|---|---|---|
| Atom composition | 0.459 | 0.875 | 0.966 |
| Edge endpoint atom-type pairs | 0.469 | 0.815 | 0.955 |
| WL structural patterns | 0.505 | 0.732 | 0.938 |
| Laplacian spectrum | 0.526 | 0.716 | 0.961 |
Results: Across all four feature sets, BNFlow pseudo graphs yield higher AUC than the normal test reference, showing that they remain distinguishable from normal graphs. Their AUC is consistently lower than that of unguided pseudo graphs, indicating that guidance brings generated samples closer to normal graphs and makes them harder to distinguish under these feature-based scores.
Normal-mode Coverage: We likewise used trained BNFlow models to generate pseudo graphs and grouped normal training graphs into three or five clusters according to the similarity of their representations. We assigned each generated pseudo graph to its nearest cluster center and compared sampling with and without guidance for each model. Normalized entropy measures dispersion across clusters, taking zero when all generated samples occupy one cluster and one when they are evenly distributed. JSD measures the difference between normal and generated cluster proportions, with lower values indicating closer agreement. The occupied-cluster percentage measures the fraction of normal clusters receiving at least one generated sample. Results are averaged over three repeated runs with different random seeds.
| Clusters | Guidance | Normalized entropy ↑ | JSD ↓ | Occupied clusters (%) ↑ |
|---|---|---|---|---|
| 3 | Off | 0.474 | 0.159 | 66.67 |
| 3 | On | 0.599 | 0.077 | 77.78 |
| 5 | Off | 0.205 | 0.294 | 46.67 |
| 5 | On | 0.342 | 0.237 | 66.67 |
Results: We see that the guidance increases normalized entropy and occupied-cluster coverage and reduces JSD for both cluster counts. These results indicate that although the generated samples do not cover all normal modes represented by these clusters, they cover a majority of the normal clusters on average.
Anomaly-type Detection: We used 73 normal graphs from the COX2 test set to construct three types of anomalies: structural anomalies by selecting a subset of edges and reconnecting one endpoint to another node, attribute anomalies by selecting nodes and randomly replacing each node's category with one of the other 34 categories, and mixed anomalies by combining edge rewiring and node-category replacement. Both variants were evaluated on identical sets of 73 normal and 73 perturbed graphs at each edge/node perturbation rate (5%, 10%, and 15%), with AUROC averaged across rates.
| Perturbation type | Without guidance AUROC (%) | BNFlow AUROC (%) |
|---|---|---|
| Structural | 67.88 | 85.87 |
| Attribute | 70.48 | 77.29 |
| Mixed | 69.78 | 84.55 |
Results: The proposed guidance improves mean detection AUROC over the unguided variant for structural, attribute, and mixed anomalies.
Response to Reviewer iCKC (4/4)
In the following, we clarify weakness 4 and question 3 concerning statistical reliability, evaluation protocols, and computational budgets.
4. Statistical Reliability, Evaluation Protocol & Computational Budgets. (W4, Q3)
4.1 Paired Significance Tests. (W4-part1, Q3-part1)
We added 10 MUTAG runs paired by identical splits and corresponding random seeds, under the same evaluation protocol. Given the limited time, we selected MUTAG because AGDiff has a short runtime on this dataset and its reported performance gap with BNFlow is relatively small.
| Statistic | Result |
|---|---|
| BNFlow mean AUROC (%) | 95.63 |
| AGDiff mean AUROC (%) | 94.31 |
| Mean paired improvement | +1.32 percentage points |
| Pairs favoring BNFlow | 8 / 10 |
| Paired t-test, two-sided | 0.0216 |
| Exact sign-flip test, two-sided | 0.0234 |
| Exact Wilcoxon signed-rank test, two-sided | 0.0273 |
Results: All tests support the MUTAG improvement at the 0.05 level.
Evaluation Protocol: All methods use the same data splits and evaluation protocol. Training uses 80% of normal graphs; testing combines the remainder with at most an equal number of anomalies on moderate-scale datasets, or all anomalies on large datasets. The paired runs share actual graph indices.
4.2 Complete Cost Accounting. (W4-part2, Q3-part2)
We use the same outer timing protocol for all baselines, covering the complete experiment subprocess, including data preparation, training, evaluation, and result saving. However, we found that the original BNFlow runs loaded cached VGAE checkpoints, whereas AGDiff's timed runs included VGAE pretraining. We apologize for overlooking this difference, which inadvertently excluded BNFlow's VGAE-pretraining cost from the original 230–800× comparison. We have now included this cost, remeasured BNFlow's total runtime, and listed the time spent in each stage below:
| Dataset | BNFlow VGAE (s) | FM pretraining + generation (s) | Joint optimization + other (s) | BNFlow total (s) | AGDiff total (s) | Ratio |
|---|---|---|---|---|---|---|
| MOLT-4 | 1,379.05 | 110.55 | 200.72 | 1,690.32 | 55,900.25 | 33.07× |
| MCF-7 | 633.70 | 189.66 | 154.58 | 977.94 | 34,238.78 | 35.01× |
| PC-3 | 830.38 | 63.01 | 55.98 | 949.37 | 35,592.08 | 37.49× |
| SW-620 | 1,502.55 | 105.90 | 199.96 | 1,808.41 | 47,677.86 | 26.36× |
Results: With VGAE pretraining included, BNFlow achieves a 26.36–37.49× speedup over AGDiff across the four evaluated datasets. The remaining runtime gap is not due to omitted VGAE pretraining, as both methods pretrain the VGAE for 100 epochs with comparable runtimes. The difference arises during joint training. During joint training, AGDiff takes roughly one hundred to over two hundred seconds per epoch across these four large-scale datasets, whereas BNFlow takes less than or approximately 10 seconds per epoch on each. A major reason is that AGDiff performs 1,000 diffusion sampling steps while BNFlow uses only 8 Euler updates, and this generation process must be executed for every batch.
Performance under an Equal 30-minute Training Budget: We additionally compared BNFlow and AGDiff under the same 30-minute (1,800-second) training budget, including VGAE pretraining for both methods, with final testing timed separately. AGDiff stopped at the budget limit before completing its planned joint training, completing fewer than 10 joint-training epochs on each of the four datasets. Within the same 30-minute budget, BNFlow pretrained the VGAE from scratch and completed its prescribed FM pretraining and joint training. AUROC and F1-score (%) are reported below.
| Dataset | BNFlow AUROC (%) | BNFlow F1-score (%) | AGDiff AUROC (%) | AGDiff F1-score (%) |
|---|---|---|---|---|
| MCF-7 | 91.63 | 90.52 | 84.91 | 84.32 |
| MOLT-4 | 88.78 | 87.78 | 79.44 | 82.76 |
| PC-3 | 92.55 | 93.41 | 12.47 | 70.66 |
| SW-620 | 94.54 | 94.43 | 16.85 | 69.73 |
Results: Under the same 30-minute training budget, BNFlow achieves higher AUROC and F1-score than AGDiff on all four datasets.
Finally, thank you again for your time and constructive comments. We hope these responses address your concerns. We would be happy to discuss any remaining questions and provide further clarification.
Response to Reviewer iCKC (1/4)
We thank the reviewer for the thoughtful and constructive comments. We address each point below.
1. Technical Distinction & Controlled Comparisons of Guidance Mechanisms. (W1, Q1)
1.1 Technical Distinction. (W1, Q1-part1)
Thank you for this important question. Below, we explain how BNFlow differs from existing pseudo-anomaly generation methods.
Gap in Existing Methods: Existing classifier-guided generation methods for GAD [1] use a shared classification objective to adjust generation through detector feedback. However, this objective can favor generating negatives that are easier to identify and does not guarantee that they remain challenging for an evolving detector. Such easy negatives may provide insufficient pressure to learn a decision boundary that closely follows the normal data distribution, limiting sensitivity to subtle anomalies. Adversarial approaches instead use a min–max objective to generate negatives that challenge the discriminator. However, SDM explicitly identifies min–max training instability as a limitation of its adversarial variants [2]. Neither a shared classification objective nor an adversarial objective alone provides fine-grained control over how much each node is refined at each intermediate generation state or how this local refinement budget adapts to the evolving detector.
Our Work Contribution: BNFlow addresses the need for fine-grained, detector-adaptive refinement by using the detector gradient to set each node's guidance direction and the learned transport velocity to bound its magnitude. Ranking feedback adjusts guidance strength as the detector evolves, while continued FM regression maintains a normal reference. This couples local difficulty control with transport-relative constraints, with Proposition 1 establishing the guidance's first-order optimality under the node-wise budget.
1.2 Controlled Guidance Comparisons. (W1, Q1-part2)
Controlled Comparison: We compared alternative generation and guidance strategies inside the same FM framework, keeping the data splits, VGAE settings, networks, and training budgets fixed. We replaced BNFlow's guided generation with joint training alone, corresponding to Without discriminator guidance (w/o DG) in Figure 5, and with SDM-style adversarial training [2]. We additionally adapted two guidance mechanisms from diffusion and flow-based generative models. DPS-style guidance [3] backpropagates through the predicted endpoint, whereas cov-A-style guidance [4] applies endpoint gradients directly with 1−t scaling. Values are AUROC and F1-score (%), mean ± standard deviation.
Results: BNFlow achieves higher mean AUROC and F1-score with lower variability on both datasets, supporting the contribution of the proposed guidance strategy beyond FM alone.
References
[1] Cai, J., et al. (2025). Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection. ICML.
[2] Cai, J., et al. (2025). Self-Discriminative Modeling for Anomalous Graph Detection. ICML.
[3] Chung, H., et al. (2023). Diffusion Posterior Sampling for General Noisy Inverse Problems. ICLR.
[4] Feng, R., et al. (2025). On the Guidance of Flow Matching. ICML.