acceptodds
Under review as a conference paper at ICLR 2027

On the Effectiveness of Resampling for Supervised Graph Anomaly Detection

Abstract

When does resampling help supervised graph anomaly detection? This empirical study compares resampling across 17 benchmarks and two tree-ensemble learners, holding graph-derived features and labelled splits fixed. We use allocation, transformation, and the weight rule to distinguish where resampling gains arise. The clearest gains occur on injected anomalies: simple feature perturbation improves test AUPRC over class weighting by 5.22 points with XGBoost and 4.40 with Random Forest; its smaller organic gains have confidence intervals including zero. Under scarce labels, labelled-split variation exceeds strategy variation in the organic panel: a descriptive decomposition assigns 72% versus 10% to these factors for XGBoost, and 68% versus 16% for Random Forest. Adaptive controls establish a small gain from residual-guided allocation and weighting, while the additional benefit of structural transformation remains unresolved. These findings motivate separating anomaly provenance, repeating paired labelled splits, and testing transformations against unchanged copies with matched allocation and class mass.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Anomaly Detection, Supervised Learning, Benchmarking, Resampling, Class Imbalance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ns8hOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes an empirical study on graph anomaly detection, which investigates the resampling technique for GAD. This paper claims several findings from an engineering perspective as its main findings, supported by evidence from evaluation metrics.

Critical strengths
  1. Analysing resampling in GAD is definitely an interesting and novel perspective in this field. I believe the topic is interesting in the supervised GAD and generalist GAD areas.
Critical weaknesses
  1. The background is unclear. The background and motivation of this study are not well explained; usually, they are supported with only a few sentences. This makes it difficult for readers to understand why this is necessary and the importance of studying resampling.

  2. This is not up-to-date research. In the related work section, it cites only 3 GAD methods. Notably, the whole related work section only cites two papers published after 2025.

  3. Although the topic itself is interesting, this report does not contribute any meaningful new knowledge or insights to the GAD community. First, the investigated baselines (XGBoost and Random Forest) are peripheral and outdated in GAD research, where the mainstream focus has shifted toward graph-learning-based algorithms for many years. If these analysed techniques do not attract any research interest, why should a paper focus on conducting empirical analysis based on these dead methods? Moreover, the report is totally focused on an engineering perspective and enumerating numbers without insights. The close-up analyses are hand-waving, rather than raising any grounded recommendations or talking about something that is relevant to the GAD itself. I suggest the authors do more background research to ensure they understand what the current practices are first, before conducting any analysis.

Questions for authors

Can you re-conduct the whole analysis on recent supervised GAD baselines?

1 comment

Thank you for the review. We agree with W1, W2 and the question. We disagree with W3.

W1 - Background and motivation are unclear. Agreed. §1 is rewritten. With about twenty labelled anomalies, the cheapest lever is the training set, and each resampling method is reported with its own detector, representation and split. Two gaps follow. A result obtained in one detector family (GNNs, tree ensembles, tabular foundation models) may not hold in another. And nobody had examined whether the label-free traces of injected benchmarks (degree and feature norm; Huang et al., ICDE 2023) drive the gains of supervised resampling.

W2 - Not up to date. Agreed. §2 is rewritten and broaden with tabular foundation models and recent GNNs litterature. It now covers GHRN, PMP, PC-GNN, ConsisGAD, SpaceGNN, APF, SAGAD, ChiGAD, TFM4GAD and LightGAD, recent work on imbalance (IGL-Bench, BAT, two 2026 studies) and the work on injected-anomaly leakage. The bibliography goes from 25 to 46 entries, and from 3 to 16 dated 2024-26.

W3 - XGBoost and Random Forest are "dead methods". This is incorrect.

  • XGBGraph is a baseline of SpaceGNN (ICLR 2025), ChiGAD (KDD 2025), APF (ICLR 2026) and SAGAD (WWW 2026).
  • TFM4GAD (WWW 2026) is a tabular model on a flattened graph. It reports a higher mean AUPRC than SpaceGNN and ConsisGAD on four GADBench datasets.
  • On our splits, with no extensive hyperparameter model tuning, the mean AUPRC over the 17 benchmarks is 39.0 for TFM4GAD, 33.2 for RFGraph and at most 29.9 for each of six GNN detectors (BWGNN, GHRN, SpaceGNN, GraphSAGE, PC-GNN, ConsisGAD). The margin is resolved against all six for TFM4GAD and against three for RFGraph (Tables 12, 19).
  • These GNN numbers are not weak runs: BWGNN, GHRN and SpaceGNN are within 0.6 points of the published means at this label budget (App. D.8).

We do not claim that trees win everywhere. The best GNN is above tuned RFGraph on five of the nine injected graphs (§4.4).

W3 - No insight, nothing relevant to GAD itself, no grounded recommendations. We disagree, three results concern GAD ecosystem directly.

  • On the standard injected benchmarks, the feature norm alone, with no label and no training, scores 25.6 AUPRC. No GNN with default/published hyperparameters we ran exceeds 22.8 on average (Table 12).
  • The gain of Gaussian-noise oversampling grows with this artefact and is not resolved without it (§4.3, Table 4).
  • Tuning hyperparameter the tree detectors adds about three points, as much as the best sampler (§4.4).

§6 turns these into recommendations: tune the detector first, repeat paired labelled splits, report organic and injected anomalies separately and by type, and check the label-free degree and norm baselines.

Q - Re-conduct the analysis on recent supervised GAD baselines.

  • Samplers now run inside GHRN, SpaceGNN, BWGNN, GATv2, TFM4GAD and TabPFN-v2. ConsisGAD, PC-GNN and GraphSAGE run as reference detectors on the same splits (Tables 2, 12).
  • The organic null holds in every family. The injected gain of the submission (+5.2 with XGBGraph) does not transfer to any GNN: between -0.3 and +0.7, none resolved.
  • Limits (§5): no sampler was inserted into ConsisGAD or PC-GNN, whose own augmentation or neighbour sampling would confound it. SpaceGNN and TFM4GAD run outside their published configuration on the injected benchmarks. The GNNs hyperparameter were not specifically tuned for each dataset.
Reviewer_nTnzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies when resampling improves supervised graph anomaly detection. It separates resampling into allocation, transformation, and weight mechanisms, and evaluates several strategies on 17 datasets with two tree-based classifiers. Results show clearer gains on injected anomalies, while organic results are strongly affected by labeled-data splits.

Critical strengths
  1. The evaluation covers 17 datasets, two learners, and repeated paired splits. Separating organic and injected anomalies and reporting confidence intervals support a careful assessment of resampling effects.

  2. The allocation-transformation-weight framework helps identify whether gains arise from sample selection, representation changes, or class weight. This provides useful guidance for evaluating future resampling methods.

Critical weaknesses
  1. The paper attributes stronger ROSE gains to injected anomalies, but the organic and injected panels use different graphs, and all injections follow the same recipe. A same-graph or multiple-injection comparison would better separate anomaly provenance from graph and injection effects.

  2. The controlled study focuses on tree ensembles with fixed graph-derived features, while the adaptive controls are mainly evaluated with XGBoost. Extending these experiments to an end-to-end GNN would strengthen the generality of the conclusions.

  3. The label-budget study fixes 80 normal labels while varying anomaly labels, changing both supervision and class imbalance. An additional experiment with a fixed class ratio would better isolate the effect of label quantity.

Questions for authors

1.Could the authors separately report results for contextual and structural injected anomalies, preferably under matched graph supports? This would clarify whether ROSE benefits from injected anomalies generally or from its alignment with the specific feature-based injection process.

2 comments

Thank you for your review; your requests enriched the study.

W1. Different graphs per panel, single injection recipe. Both are now tested (Sec. 4.3).

  • Same graph: with the standard recipe injected into five organic graphs, the injected-minus-organic difference in ROSE's gain is +0.6 [−2.7, +4.9] for XGBGraph and −6.0 [−10.0, −1.7] for RFGraph (Table 27). The panel contrast is not reproduced, so we no longer attribute the gain to injection as such. With five graphs, we do not claim the opposite either.
  • Several recipes (nine graphs; Tables 4, 23 to 25): with one candidate per swap ROSE gains nothing. The gain is resolved from 15 candidates and reaches +6.6 (XGBGraph) and +10.4 (RFGraph) at 50, while the norm signal grows from 1.1 to 13 times prevalence. A norm- and degree-preserving recipe gives no resolved gain.
  • Limit: removing the norm signal also makes the swaps hard to detect, so the two are not separated.

W2. Tree-centric study; adaptive controls with XGBoost only.

  • End-to-end GNNs: samplers now run inside BWGNN, GATv2, GHRN and SpaceGNN, on the hidden representation and, for ROSE, also at the input (Table 2, App. B.3). No sampler has a resolved gain on either panel. Heads retrained on frozen embeddings give one resolved gain in twelve contrasts (Table 22).
  • Second learner: with RFGraph, views against copies remain unresolved in both panels (Tables 5, 28). Every RFGraph estimate is positive and one of two designs is resolved over all 17 benchmarks (+0.44), so we write "no resolved benefit within either panel", not "no benefit".
  • Limit: these controls rely on boosting residuals and have no end-to-end counterpart (Sec. 5).

W3. Fixing 80 normals changes supervision and imbalance together. We crossed 5, 10 and 20 labelled anomalies with 20, 40 and 80 labelled normals on the same six benchmarks (Sec. 4.5, Table 13).

  • Fixed 1:4 ratio, four times the labels (contrast fixed in advance): ROSE's gain over class weighting does not change, −2.3 [−8.8, +3.3] for XGBGraph and +0.5 [−3.0, +4.0] for RFGraph.
  • More anomalies, fixed normals: on injected benchmarks the gain falls, resolved in every cell with five anomalies and in none with twenty. On organic benchmarks nothing is resolved. With six benchmarks we cannot separate the number of anomalies from the ratio, and the paper no longer reads the original curve as a label-quantity effect.

Q1. Contextual and structural anomalies separately, on matched supports. Done for all ten configurations (Fig. 1, Table 17). On attribute swaps the gain is resolved in all six tabular configurations (+2.6 to +10.8) and in none of the four GNNs. On cliques no configuration gains. On matched supports (Table 27), the swap gain replicates on organic hosts (+3.0 and +5.2, both resolved) and cliques do not gain. This is your hypothesis: ROSE benefits from its alignment with the feature-based injection, not from injected anomalies in general.

Thank you for the author's response. It has partially addressed my concerns, and I have also read the comments from the other reviewers. I will maintain my score.

Reviewer_8jUDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an empirical study of resampling for supervised graph anomaly detection under limited labels. It evaluates multiple strategies across 17 benchmarks using XGBoost and Random Forest, with controlled comparisons of sample allocation, representation transformation, and weighting. The results show clear benefits on injected anomalies but less conclusive gains on organic anomalies.

Critical strengths
  1. The paper addresses a practical and relevant question: whether resampling improves graph anomaly detection when labelled anomalies are scarce.
  2. The allocation, transformation, and weighting framework provides a clear way to disentangle the effects of different resampling mechanisms.
  3. The evaluation is comprehensive, covering 17 datasets, two learners, repeated labelled splits, and different supervision budgets. Comparisons with both weighted and unweighted baselines strengthen the analysis.
Critical weaknesses
  1. The main contribution is empirical rather than methodological, and the novelty may be limited for ICLR. The controlled comparisons also rely on fixed graph-derived features and tree ensembles, leaving generalization to graph neural networks unclear.
  2. The differences between organic and injected anomalies are not sufficiently explained. Different graph settings and a limited anomaly injection protocol make it difficult to determine which data characteristics drive the observed gains.
  3. The benefits of adaptive resampling remain modest. Residual-guided improvements are small and configuration-dependent, while simple perturbation often performs better. The evidence for a substantial advantage over simpler resampling strategies is therefore limited.
Questions for authors

See weaknesses.

1 comment

Thank you very much for the review. We adressed your flagged weaknesses.

W1 - Empirical contribution; fixed features and tree ensembles leave generalization to GNNs unclear.

  • Generalization, tested. The comparison now covers ten detector configurations in three families, including four GNNs trained end to end with samplers on the hidden representation (Table 2). The injected gain doesn't generalize: in the four GNNs ROSE's gain is between -0.3 and +0.7, none resolved, and placing the noise at the input does not change this (§4.2). Two findings do generalize: the organic null (4.1), and the weight of the labelled split, which accounts for 53-79% of AUPRC variation against 3-16% for the strategy across the ten configurations (Table 18).
  • Nature of the contribution. The paper stays indeed empirical. What it contributes is now stated as such: an evaluation result about the standard injected benchmarks (4.3), a control that separates what a sampler selects, generates and reweights (4.6), and evidence that tuning the detector is worth as much as the best sampler (4.4).

W2 - The organic/injected difference is not sufficiently explained. A section of the paper is now dedicated to this question (4.3). We describe where the gain comes from, and we state what remains open.

  • It is tied to one anomaly type. Attribute swaps gain in all six tabular configurations; cliques gain in none (Fig. 1, Table 17).
  • It tracks a label-free artefact of the recipe. The standard recipe copies the most distant of 50 candidates, which shifts the feature norm of swapped nodes. With fewer candidates the norm signal and ROSE's gain shrink together, down to zero with one candidate; under a norm-preserving swap no gain is resolved (Table 4).
  • The detector can be given the signal directly. Appending degree and feature norm to the features raises class-weighted XGBGraph by +8.1 points on the injected panel, more than ROSE, and halves ROSE's gain for that detector. RFGraph barely uses the two columns, so this holds for one learner (Table 21).
  • It is not provenance. With the same recipe injected into five organic graphs, the difference between injected and organic tasks is not reproduced (Table 27). We withdrew that reading.
  • What remains open. Removing the norm signal also makes the swaps hard to detect, so the two are not separated. The leak is not sufficient either: Cora, Citeseer and Pubmed leak 14 to 25 times prevalence and gain nothing on swaps.

W3 - Adaptive gains are modest; simple perturbation often does better. We agree, and the revision presents this as a result (4.6, Table 5).

  • At matched allocation and class mass, graph views are not distinguishable from plain copies of the same nodes, in either panel, with XGBoost and now with RFGraph.
  • Residual allocation gives XGBGraph +0.2 to +0.3 points and RFGraph nothing resolved nor utterly significant
  • The simple perturbation that did better is itself narrow: tabular detectors, attribute swaps, and a recipe artefact (4.2–4.3)

Comments on the paper

1 comment
AuthorsOpenReview ↗

General answer to reviewers

First and foremost we thank the reviewers for their time. Here's a summary of what changed :

SubmissionRevision
Detectors that receive samplersTrees methods adapted to graph learning10 configurations in 3 families: Tabular foundation models (in-context); trees methods adapted to graph learning; GNNs (Table 1)
Injectionone recipe, on graphs other than the organic onescandidate ladder, a norm- and degree-preserving recipe, and the standard recipe injected into organic graphs (4.3)
Injected anomalies by typenot reportedattribute swaps and cliques, for every detector (Fig. 1, Table 17)
Label budgetanomalies varied at 80 normals(k, n) grid (4.5)
Bibliographyexpanded on the most recent methods

The emphasis on full supervision was reduced.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.