acceptodds
Under review as a conference paper at ICLR 2027

HADES: Self-Validated Energy Selection for Structural Out-of-Distribution Detection on Hypergraphs

Abstract

Hypergraphs represent higher order interactions among groups of entities, and hypergraph neural networks (HNNs) exploit these relations for prediction. At deployment, hyperedge formation may change, placing nodes in relational contexts unseen during training. Structural out of distribution (OOD) detection flags when confident predictions rely on relational patterns no longer supported by training, making it essential for reliable deployment. Yet this task remains underexplored, and no fixed post hoc score is reliable across datasets and backbones. We introduce HADES (Hypergraph Adaptive Detection via Energy Selection), which pairs a new structural score with adaptive detector selection. The score measures disagreement between log partition confidence propagated over the current incidence and a soft incidence anchored by prototypes fitted to clean training embeddings. To handle detector heterogeneity, HADES turns score choice into a predeployment self validation problem. Matched perturbations of the clean hypergraph define a surrogate shift. Using this surrogate, HADES selects and freezes the single scoring rule best aligned with anticipated deployment changes. Across 12 hypergraphs and four backbones, HADES achieves the best average detector rank on every backbone. Its overall average rank is , compared with for runner up KLM. These results show that reliable structural OOD detection depends on both a clean relational reference and score selection aligned with the anticipated shift.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
HNN, OOD, Adaptive, Energy, Structure

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_n67nOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents HADES, a node-level structural OOD detection method for hypergraphs. HADES introduces a prototype-based energy score that compares confidence propagation between the current hypergraph and a reference built from clean training representations (Lines 176 - 178). Moreover, it generates synthetic structural shifts on validation data to evaluate eight candidate detectors, then selecting the best-performing one for detection. The paper presents a practical approach to structural OOD detection with diverse experiments. However, I have some concerns with the method and experiments, particularly on how the improvements appear to rely heavily on selecting existing detectors using synthetic validation shifts, and whether this method generalizes to unknown shifts.

Critical strengths
  1. The overall storytelling is clear and generally easy to follow.

  2. I really enjoyed the presented figures, particularly HADES overview.

  3. The paper evaluates HADES across multiple datasets and compares with several baselines, showing improvements in detection metrics like AUROC and FPR95.

  4. HADES also investigates an important problem of structural OOD detection in hypergraphs, a relatively underexplored setting compared with traditional graph OOD detection.

Critical weaknesses
  1. The main experiments use the same perturbation approach and settings for generating validation and test OOD samples (section 4.1), which could potentially make detector selection easier. Although the authors explored different perturbation severities and natural structural shifts in the appendix, their results reveal potential failures in some settings. For example, HADES achieves only 37.7 AUROC on the cocitation --> coauthorship transfer, compared with 70.4 for ViM (Table 7 in appendix). This raises concerns about the reliability of detector selection when the actual structural shifts differ from those used during validation.

  2. Also, a single fixed train/validation/test split was used for each dataset (Line 287), while results were reported over three random seeds (Line 335). Since the training and validation data directly affect prototype construction and detector selection, it is unclear how sensitive the results are to different data splits.

  3. Although the paper analyses the theoretical complexity, it does not provide empirical runtime or memory comparisons. Given that HADES requires evaluating numerous detectors before selecting one, clearly reporting the additional computational cost maybe beneficial to demonstrate the efficacy.

  4. While the prototypes are learned from clean training data, the prototype assignments are computed using embeddings from the potentially shifted hypergraph. As a result, both the current and reference views may be affected by structural changes. It is unclear whether their difference can reliably identify structural OOD samples.

  5. Based on Alg 1, HADES selects detectors solely based on validation AUROC. It would be interesting to explore whether using alternative metrics, such as FPR95, which is more sensitive and constrained, could lead to better OOD detection performance.

  6. General note: It is recommended to review and discuss more recent graph OOD detection papers in the extended related work section, as the current discussion focuses primarily on works published before 2024. E.g.,

[1] Fan, X, et al., HGOOD: Hypergraph-enhanced Graph Contrastive Learning for Graph Out-of-Distribution Detection. IJCAI26

[2] Wang D, et al., What Information Matters? Graph Out-of-Distribution Detection via Tri-Component Information Decomposition. ICML26.

[3] Wang, Y, et al., CLINIC: Towards High-quality Graph Out-Of-Distribution Detection. ICML26

Questions for authors
  1. Why does HADES select only one detector instead of combining multiple detectors? Could the authors compare their approach with simple averaging, weighted averaging, or other detector combination methods? Do the authors anticipate any benefits with this approach?

  2. How sensitive is HADES to different train/validation/test splits?

  3. The ablation results suggest that most improvements come from detector selection rather than the proposed discrepancy score. Could the authors further clarify the contribution of the new score and compare against selecting from a larger pool of existing detectors?

  4. How does HADES perform under shifts like feature or label shifts, that are also commonly explored in graph OOD studies?

5 comments
AuthorsOpenReview ↗

We address detector combination, split sensitivity, the contribution of the discrepancy score, feature shifts, and the three suggested methods.

Q1. Single detector or combination

We thank the reviewer for suggesting combinations. To compare them on equal terms, we used the eight HADES candidates, clean validation nodes, and surrogate copies. We standardized candidate scores on clean validation nodes. Averaging weights came from surrogate AUROC, while logistic regression learned from proxy labels on the surrogate task.

Table R1. Mean AUROC (%) across dataset and backbone settings.

RulePairedMixed nodes, q = 0.3
Average all eight60.264.3
Average top two by surrogate AUROC68.175.2
Softmax weights from surrogate AUROC68.276.4
Signed weights from surrogate AUROC71.375.5
HADES, select one67.976.4
Logistic combination trained on surrogates77.579.1

Equal averaging dilutes strong candidates when other scores respond in the opposite direction. Top two averaging and softmax weighting nearly match single selection. In contrast, the learned combination improves AUROC by 9.6 points in the paired protocol and 2.7 points on mixed nodes. The same surrogate validation task supports stronger fusion without target OOD labels. We retain single selection as the submitted method because it scores only one detector at deployment and has the regret guarantee in Proposition 2. We report learned fusion as a separate variant.

Q2. Sensitivity to data splits

We thank the reviewer for raising split sensitivity. The submission uses one fixed train, validation, and test split with three master seeds. Across those seeds, HADES has a mean AUROC standard deviation of 2.3 points, compared with 2.4 to 4.6 for other detectors, and its mean rank stays between 2.5 and 2.7. We also tried a new random partition of nodes outside the test set into reference and validation sets. On mixed nodes at q = 0.3, this variant reaches 75.7 AUROC versus 76.4 with the submitted partition. These checks address training randomness and one reference partition. They do not measure sensitivity to independently redrawn train and test splits, which would require retraining the backbones.

Q3. Discrepancy score and a larger pool

We thank the reviewer for separating the score from selection. We removed the discrepancy score from both the submitted pool and an expanded pool of existing post hoc detectors, while keeping the selection procedure fixed.

Table R2. Selected AUROC (%) in the paired protocol, mean over 48 settings.

Candidate poolWith sdiscs_{\mathrm{disc}}Without sdiscs_{\mathrm{disc}}Gain
Submitted eight candidates67.865.02.8
Up to 17 existing scores plus sdiscs_{\mathrm{disc}}71.670.11.5

A larger pool alone raises AUROC from 65.0 to 70.1. The discrepancy score then adds a further 1.5 points, with improvement in nine settings and a decline in one. In the submitted pool it adds 2.8 points, with improvement in 15 settings and a decline in two. Selection is a major source of the gain, while the score supplies complementary evidence from the difference between current incidence and prototype anchored views.

We also examined the three suggested papers. HGOOD[1] and CLINIC [3] detect OOD graphs among graph samples. Their reported graph level results therefore do not directly match node level detection in one hypergraph. TIDE [2] is closer because it detects OOD nodes under feature and structure shifts. It trains a new graph model with an information bottleneck, whereas HADES evaluates scores on a frozen hypergraph backbone. These distinctions explain why we discuss them rather than insert them into the post hoc score pool.

Q4. Feature and label shifts

We thank the reviewer for asking about other shift types. We evaluated the feature interpolation of Appendix C.2 at strengths 0.1 and 0.5, applied either to all nodes or to 30% of nodes in one hypergraph. The submitted selector used structural surrogates. The second HADES column used feature surrogates but the same detector pool and validation rule.

Table R3. Feature shift AUROC (%), mean over 48 settings.

Strength and affected nodesStructural surrogatesFeature surrogatesBest fixed detector
0.1, all53.656.557.1
0.1, 30%50.051.051.3
0.5, all63.672.166.5
0.5, 30%55.058.257.7

At strength 0.1, all methods are near chance. At strength 0.5, feature surrogates raise HADES from 63.6 to 72.1 on all nodes and from 55.0 to 58.2 on mixed nodes. This shows that the validation rule can transfer to a feature shift when its surrogate reflects that shift. Unseen class label shift remains untested because a class holdout changes the backbone training set and requires retraining.

I appreciate the authors' response and additional experiments to address my concerns. However, several important issues remain:

From my understanding, the Q4 response further shows that HADES relies heavily on how closely the surrogate validation shifts has to resemble the actual OOD shifts in order to improve performance. For example, changing from structural to feature surrogates improves AUROC from 63.6 to 72.1. However, in practice, the distribution of the test OOD shifts are unknown, making it difficult to ensure such alignment. So this raises fundamental concern for me about the robustness and practical applicability of HADES under unseen distribution shifts.

Additionally, the learned detector combination substantially outperforms HADES's single-detector selection (77.5 versus 67.9 AUROC) using the same candidates and surrogate data. Although the authors stated that single selection is computationally cheaper, no empirical runtime or memory comparisons are provided to justify this performance trade-off.

Also, based on the Q3 results, i am hesitant about how expanding the pool of existing detectors improves AUROC from 65.0 to 70.1, while adding the proposed discrepancy score provides only a further 1.5 point improvement. This suggests that most of the performance gain comes from selecting existing detectors rather than the proposed score.

Therefore, given my concerns still remain and several of the weaknesses are not addressed (e.g., W2, W3, W5). I would like to maintain my current rating.

AuthorsOpenReview ↗

We provide additional evidence on unseen shifts, runtime, prototype stability, and FPR95 selection, while clarifying data splits and recent related work.

Q1. Shifts unseen during validation

We appreciate the concrete Cora transfer example. We tested a stricter setting in which the target shift type is excluded from the surrogate library. For each of eight target types, validation uses only the other seven types. The natural transfer between cocitation and coauthorship has no corresponding surrogate.

Table R4. AUROC (%). Synthetic: 48 settings per type. Natural: 24 settings.

RuleMean across eight typesWeakest typeNatural transfer
Best fixed detector for each column59.155.351.6
Submitted single selection62.155.058.6
Single selection, target type excluded62.955.463.9
Learned combination, target type excluded67.056.077.2

The submitted selector exceeds the best fixed detector on average, and withholding the target type does not erase that advantage. Combining candidates trained on the other types raises the mean by another 4.1 points. For the cited HGNN cocitation to coauthorship case, AUROC rises from 37.7 to 52.9 with combination (ViM: 70.4). Across all four backbones in that direction, combination averages 71.3 versus 51.8 for the best fixed detector. This fusion uses the same surrogate validation framework and candidate pool as single selection.

Q2. Data splits

All three master seeds use the same data partition. Across them, HADES has 2.3 AUROC points of mean standard deviation, versus 2.4 to 4.6 for the other detectors. Repartitioning nodes outside the test set once gives 75.7 AUROC on mixed nodes, compared with 76.4 under the submitted reference and validation partition. These checks support stability to training randomness and this validation partition. We will evaluate new data splits with retrained backbones in the revision.

Q3. Runtime and memory

We profiled self validation and target scoring over 48 settings. Self validation takes 1.13 seconds and uses 644 MB of additional peak GPU memory on average per trained backbone. The median fixed detector takes 0.013 seconds to score a target hypergraph.

Table R5. Deployment costs, seed 42. GPU memory is above the resident state.

RuleMean timeMean peak GPU memory
Submitted single selection0.019 s198 MB
Learned combination0.13 s323 MB

The combination evaluates eight candidates sequentially. Its time sums their separately measured scoring times, and its memory uses the largest measured candidate peak. Logistic fitting takes a few hundredths of a second and reuses the surrogate scores computed during self validation.

Q4. Prototype view under shift

Fixed centers constrain the prototype view even though shifted embeddings change node assignments. Across settings, its drift is 0.52 to 0.76 of the current incidence view drift. Removing the discrepancy score lowers the learned combination by 3.3 AUROC points with matched surrogates and 1.5 points when the target type is excluded. Thus the view difference contributes structural evidence even when assignments move.

Q5. Selection metric

We compared selecting the single detector by surrogate AUROC and by surrogate FPR95. The latter improves target FPR95 in both validation settings.

Table R6. Target FPR95 (%), lower is better. Mean over eight shift types.

Selection criterionMatched typeTarget type excluded
Surrogate AUROC79.882.5
Surrogate FPR9577.081.5

The submitted AUROC criterion remains best for target AUROC. The FPR95 criterion lowers target FPR95 by 2.8 points with matched surrogates and 1.0 point when the shift type is excluded. We will make the criterion configurable for the intended operating metric.

Q6. Recent graph OOD work

We agree that the extended discussion should better cover recent work. Our main text already discusses UniGOD, BaCa, PGOS, and HyperGOOD from 2026. We will also position HGOOD, TIDE, and CLINIC. HGOOD and CLINIC detect entire graphs. TIDE detects OOD nodes but trains a new graph model. Their published results do not directly compare with node detection using a frozen hypergraph backbone.

Reviewer_783zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HADES for node-level structural OOD detection on hypergraphs. HADES introduces a new prototype-anchored energy discrepancy score and combines it with an adaptive detector selection procedure. It fits prototype centers to clean training embeddings, then softly assigns current embeddings to these fixed centers to construct a reference incidence. The discrepancy score measures standardized disagreement between confidence propagated through the current and reference incidences. For detector selection, Shift-Aligned Self-Validation evaluates a fixed candidate pool using three structurally perturbed hypergraphs and selects the candidate with the highest average proxy AUROC for final scoring. Experiments on 12 hypergraphs with four backbones demonstrate that HADES achieves improved performance.

Critical strengths

S1. The empirical evaluation is broad. The paper compares HADES with 13 fixed detectors across 12 datasets, and the transfer experiments help explain its gains and limits.

S2. The discussion of related work is comprehensive and clear.

Critical weaknesses

W1. The definitions of ID and OOD nodes in actual deployment need to be clarified.The paper should explain what types of structural changes are considered normal or abnormal, rather than simply treating nodes under clean and perturbed incidences as ID and OOD, respectively.

W2. The information available to the detector needs to be clearly defined. A practically important setting is detection using only a pretrained backbone and the input hypergraph, without access to the original training data. HADES requires the clean hypergraph for offline prototype fitting and surrogate generation. The paper should clarify this access requirement, what information each baseline can use, and what must be retained for deployment.

W3. The novelty and deployment assumptions of Shift-Aligned Self-Validation need further justification. Selecting the best detector from a fixed candidate pool offers limited novelty, and scalability to larger pools remains unexplored. The method relies on a structural perturbation prior that matches the target perturbations in the main experiments, but the actual deployment shift may be unknown. The paper should discuss a more adaptive selection strategy.

W4. The stability of the prototype-anchored view needs more discussion. Both the current-incidence and prototype-anchored views can be affected by the same structural shift. It remains unclear whether the prototype-anchored view provides a sufficiently clean reference for their disagreement to reliably indicate structural OOD.

W5. The computational overhead of Shift-Aligned Self-Validation is not evaluated. The procedure evaluates multiple candidate detectors on several structurally perturbed hypergraphs, potentially introducing substantial additional costs. The authors should report this overhead and compare it with the computational cost of using fixed detectors.

Questions for authors

Q1. How are ID and OOD nodes defined in actual deployment? What distinguishes normal structural changes from abnormal ones?

Q2. Can HADES work with only a pretrained backbone and the input hypergraph?

Q3. What does the selection rule add beyond existing detector-selection methods? How should surrogate perturbations be chosen when deployment shifts are unknown, and how does the method scale to larger candidate pools?

Q4. Can the authors provide evidence that the prototype-anchored view remains a sufficiently clean reference under structural shifts, and that disagreement between the two views reliably indicates structural OOD?

Q5. What are the additional runtime and memory costs of Shift-Aligned Self-Validation compared with fixed detectors?

4 comments
AuthorsOpenReview ↗

We clarify deployment definitions and input requirements, and provide new evidence on surrogate selection, prototype stability, pool scaling, and computational costs.

Q1. ID and OOD at deployment

We appreciate the request for an operational definition. HADES returns one ID score per node. An application can set a threshold from available clean validation scores, but no OOD labels are observed at deployment. Benchmark labels come from provenance: all test nodes scored on the shifted copy are OOD in the paired protocol, while only test nodes with changed incidence rows are OOD in the mixed protocol. HADES then scores whether those changed relations still support the frozen classifier. For matched hyperedge additions, within class edges raise accuracy by 4.8 points and give HADES 48.9 AUROC. Class mixing edges lower accuracy by 2.0 points and give 59.8 AUROC. The distinction concerns the effect on learned relational evidence, not the number of edits.

Q2. Backbone and input hypergraph only

We thank the reviewer for asking whether the clean training graph is essential. Submitted HADES uses it for prototypes and reference scores, along with held out validation indices and training labels for Mahalanobis. Our deployment only variant splits unlabeled nodes for reference and validation, perturbs the input incidence, and uses backbone predictions as Mahalanobis pseudo labels.

Table R1. AUROC over 48 settings.

ShiftSubmittedDeployment onlyPool oracleBest fixed
Mixed nodes, q = 0.376.475.776.9NODESAFE 69.1
Whole graph67.960.469.2KLM 60.7

With local change, the deployment only version loses 0.7 points and needs no training graph or class labels. When the whole graph shifts, its reference shifts too, and the clean training graph becomes important.

Q3. Selection and surrogate choice

We appreciate the question about what the selector adds. HADES follows synthetic anomaly selection in using proxy labels. Its proxies change hyperedge incidence while retaining node identities and features. Candidates are ranked on held out validation nodes before target scoring. Table R2 compares this structural proxy with transfer from other datasets and agreement among candidate scores on target data.

Table R2. AUROC over 48 settings.

Selection rulePairedMixed q = 0.3
Pool oracle68.876.9
HADES self validation67.976.4
Best candidate on other datasets60.068.8
Model centrality on target scores58.359.3
Random candidate55.358.1

Within the mixed protocol, a moderate severity mismatch has little effect. Selection at q = 0.3 obtains 75.9 AUROC when q = 0.5, versus 76.1 with matched severity. Pooling unrelated shift families gives average rank 5.5, close to 5.6 from reusing the submitted choice. Choosing surrogates therefore starts with the kind of relation likely to change.

Larger pools remain feasible, but candidate quality matters. Across 200 random subpools per run, increasing pool size from 2 to 16 raises selected AUROC from 61.2 to 71.5, while oracle regret rises from 1.2 to 5.9. Candidate scoring grows roughly linearly with pool size.

Q4. Prototype view

We thank the reviewer for asking how much the reference moves under shift. Prototype centers stay fixed, but shifted embeddings change node assignments, so the reference is not invariant. We compared its drift with the current incidence view on shifted nodes.

Table R3. View behavior over 48 settings.

QuantityPairedMixed q = 0.3
Median ratio of prototype to current view drift0.760.52
Nodes with less prototype view drift62%70%
Separation condition met3.5%7.4%
Correlation of condition share with score AUROC0.630.70
Standalone discrepancy AUROC59.160.1

The view is usually more stable, but disagreement is not decisive for every node. The condition is sufficient rather than universal. Where it holds more often, the discrepancy works better, so HADES validates this score alongside other candidates.

Q5. Runtime and memory

We appreciate the request for resource costs. We profiled the 48 settings with seed 42 on one RTX 5060 Ti. Self validation includes generating three surrogate copies, fitting prototypes, and scoring eight candidates. For a fixed shift family, it is performed once per trained backbone. Deployment then applies one selected detector.

Table R4. Mean costs above the resident backbone and graph.

OperationTime (s)Peak GPU memory (MB)
Self validation1.13644
Selected detector at deployment0.019Within 1 MB of fixed detector
Fixed detector0.013328

Fixed detector time is the median across 13 scores, while its memory is their maximum. The most expensive validation run takes 4.9 seconds and reaches 5.0 GB above the resident state, versus 1.5 GB for the largest fixed score. The additional cost is concentrated before deployment.

Thanks for the clarifications and additional experiments. They’re helpful, but I still have a few main concerns that haven’t been fully addressed.

1.It remains unclear how normal structural variation is distinguished from OOD changes that undermine the trained classifier. For example, a researcher joining a new collaboration group could be a normal change. What makes such a change OOD, rather than normal variation? The added experiment reports 48.9 AUROC for within-class additions and 59.8 for class-mixing additions. This suggests that HADES struggles with the former and has only limited success with the latter. For datasets where groups normally contain nodes from different classes, how does the method help distinguish normal changes from OOD?

2.The expanded candidate pool needs more explanation. The paper uses 8 fiexed candidates, but the response tests pools of 2–16. Where do the additional candidates come from? How are the subpools sampled? Without these details, it is hard to interpret the results. My earlier question also remains: what abnormality signals does the discrepancy score capture that the existing detectors miss? The standard deviations are quite large in some experiments, such as HADES on Zoo (61.3 ± 14.3, table 1) and the discrepancy score on Cora-CA (61.3 ± 14.9, table 2) and Mushroom (65.6 ± 16.4, table 3). Are the improvements consistent across runs?

AuthorsOpenReview ↗

We clarify how HADES handles normal structural variation, explain the expanded candidate pool, and provide evidence for the discrepancy score and consistency across runs.

1. Normal structural variation

We thank the reviewer for the collaboration example. A new hyperedge is not intrinsically OOD. HADES evaluates the resulting node evidence relative to the clean regime learned by the backbone, rather than judging whether a group is class pure. On Senate and House, within hyperedge class agreement is 0.48 and 0.50, near the random rate of 0.50, so mixed groups already occur in the clean reference. In a controlled test of classifier relevance, within class additions raise accuracy by 4.8 points while class mixing additions lower it by 2.0. The fusion extension separates these cases even when class mixing is excluded from validation (Table R1). Its near chance ranking of within class additions does not imply a fixed alarm rate, which can be calibrated on clean validation scores. The additions are diagnostic probes, not universal definitions of normality.

Table R1. AUROC (%) for equal sized additions, mean over 48 settings.

DetectorClass mixingWithin class
DegreeZ structural anomaly score60.356.8
Submitted HADES selection59.848.9
HADES fusion, mixing excluded from validation67.549.9
HADES fusion, full surrogate library69.751.1

2. Candidate pool and sampling

We appreciate the request for a reproducible account of the enlarged pool. Table R2 gives all 18 post hoc scores, consisting of the submitted eight, five fixed comparators, and five additional methods. NODESAFE is excluded because it requires backbone retraining. SCALE and ASH need nonnegative activations, leaving 16 scores on AllSetTransformer and ED HNN. For each run and pool size 2, 4, 8, 12, 16, and 18 where defined, auxiliary seed 0 draws 200 subsets uniformly without replacement. This seed controls only candidate subset sampling. Every candidate is evaluated on the same clean graph and three surrogate copies. Within a subset, we select by mean surrogate AUROC on held out validation nodes, then average target AUROC over draws, master seeds, datasets, and backbones. The submitted pool is one specified set of eight, not a random subset. Selected AUROC rises from 61.2 at size 2 to 71.5 at size 16, while the full available pool reaches 71.6. Target labels never enter selection.

Table R2. Candidate composition of the expanded pool.

GroupCountCandidates
Submitted8sdiscs_{\mathrm{disc}}, GNNSafe, Energy, MSP, ODIN, kkNN, Mahalanobis, ViM
Fixed comparators5MaxLogit, GEN, KL Matching, Relative Mahalanobis, NNGuide
Additional post hoc scores5GRASP, NECO, fDBD, SCALE, ASH

3. Evidence supplied by the discrepancy score

The discrepancy score compares the same confidence propagated through current and prototype anchored relations. This can reveal a changed relational context even when node energy barely moves. We test that claim on low energy change nodes and measure both rank correlation and unique detections at 95% ID acceptance. Table R3 shows a complementary relational signal.

Table R3. Complementarity checks, mean over 48 settings.

DiagnosticResult
Half of shifted nodes with the smallest energy changesdiscs_{\mathrm{disc}}: 59.0 AUROC. GNNSafe: 56.2. Energy: 51.8
Mean rank correlation with other candidates−0.12-0.12 to 0.180.18
Shifted nodes flagged only by sdiscs_{\mathrm{disc}} at 95% ID acceptance4.7% paired, 4.0% mixed

4. Consistency across runs

The reported standard deviations warrant a run level check. Across 144 paired runs spanning datasets, backbones, and seeds, HADES is among the top three detectors in 119 and beats each fixed detector in 64% to 94% of runs. Table R4 examines the three cited datasets. Zoo has only 26 test nodes, and fixed detectors vary on the same runs. On Cora CA and Mushroom, self validation chooses other candidates when the standalone discrepancy varies, keeping the selected result steadier.

Table R4. Run level checks across three master seeds.

DatasetHADES resultComparison
Zoo, 26 test nodes61.3±14.361.3\pm14.3GEN deviation 20.8, kkNN deviation 19.1
Cora CA80.7±3.280.7\pm3.2, selects kkNN in all runsStandalone sdiscs_{\mathrm{disc}}: 61.3±14.961.3\pm14.9
Mushroom87.6±4.687.6\pm4.6, selects GNNSafe or ODINStandalone sdiscs_{\mathrm{disc}}: 65.6±16.465.6\pm16.4

Thank you for the clarifications and additional experiments, especially the evidence that the discrepancy score adds complementary information. Some of the additional results are still not clearly explained, and I do not see sufficient new insights beyond existing detector selection approaches. I therefore maintain my rating of 2.

Reviewer_KvWsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies structural OOD detection on hypergraphs, where node identities and attributes remain fixed while hyperedge incidence changes. HADES introduces a prototype-anchored score that measures disagreement between confidence propagated through the current structure and a reference structure derived from clean training prototypes. It then uses synthetic structural perturbations to select a single detector from this score and established alternatives. Evaluation across multiple datasets and hypergraph backbones examines detection performance, component contributions, and transfer across structural shifts.

Critical strengths

1.The paper studies structural OOD detection on hypergraphs, which is a relatively new and interesting problem setting.

2.The experimental evaluation is fairly comprehensive, covering 12 hypergraph datasets and 13 baselines, together with ablation studies, hyperparameter sensitivity analysis, and robustness experiments.

Critical weaknesses

1.Unclear distinction from structural anomaly detection. The operational task closely resembles detecting structural anomalies or injected corruptions in hypergraphs. The paper does not sufficiently clarify what additional detection requirement the OOD formulation introduces, leaving the distinctiveness of the task contribution uncertain.

2.The experimental setup appears closer to graph-level OOD detection than node-level OOD detection, and under this construction, propagating OOD scores to identify OOD nodes may make the task relatively easy. Since the OOD graphs preserve the original nodes and features while only modifying the relational structure, the ID and OOD samples can be viewed as two separate graphs with matched node identities and features; consequently, scores extracted by a frozen pretrained HNN backbone may naturally develop distributional differences after propagation on the perturbed graph, which also makes the setting resemble anomaly detection more than standard OOD detection.

3.The detection objective depends on a prescribed perturbation family. The main protocol selects and evaluates detectors using matched types of structural modification. This leaves uncertainty about whether the resulting performance reflects a broader ability to identify unfamiliar relational contexts or sensitivity to the specified corruption process.

4.Limited conceptual novelty of the proposed score. The score combines prototype clustering, confidence propagation, and a standardized absolute difference between two outputs. Although this construction is understandable, its conceptual advance beyond a general consistency check between relational views appears modest.

5.Although the paper compares against a relatively large number of baselines, most of them are classical or older methods, with limited coverage of more recent approaches from the past two years, making it difficult to fully assess the competitiveness of HADES. Moreover, the component analysis suggests that the selection module contributes most of the improvement over the discrepancy score alone; since the selector mainly ranks existing detectors based on anticipated perturbations, the main gain appears to come from effective scenario-specific detector selection, while the methodological advance in the underlying OOD detection mechanism seems relatively limited.

Questions for authors

See Weaknesses.

1 comment
AuthorsOpenReview ↗

We clarify the task and scope of HADES and report new node level tests, structural shift experiments, score ablations, recent baseline comparisons, and selection results.

Q1. Task distinction

We thank the reviewer for pressing on this distinction. Structural OOD asks whether a node's new relations still support its frozen classifier as the training relations did. We added equal numbers of hyperedges with matched size distributions in two ways. Class labels served only to construct this diagnostic. Within class additions raise test accuracy by 4.8 points, while class mixing additions lower it by 2.0.

Table R1. AUROC on matched additions, averaged over 48 settings and three seeds.

DetectorClass mixingWithin classDifference
DegreeZ60.356.83.5
HADES59.848.910.9

DegreeZ uses the absolute z score of log degree against clean training nodes. HADES distinguishes changes with different consequences for learned relational evidence.

Q2. Node detection within one hypergraph

We appreciate the concern about node localization. The paired protocol alone does not establish it. We therefore tested a single hypergraph containing changed and unchanged test nodes. Only sampled incidence rows can change. All test nodes are scored together, so a graph constant score obtains 50 AUROC. We ran the same selector on mixed node surrogates from clean data. Neither choice uses target OOD labels or scores.

Table R2. Mixed protocol AUROC.

Detectorq = 0.3q = 0.5
GNNSafe68.867.5
NODESAFE69.167.7
HADES, submitted selection66.863.1
HADES, mixed node surrogates76.476.1

Here q is the probability that a node joins the shifted set. With mixed node surrogates, HADES leads the strongest fixed detector by 7.3 and 8.4 points.

Q3. Perturbation scope

We thank the reviewer for probing the surrogate assumption. The gain in Q2 depends on whether the surrogates reflect the target shift. We tested four other structural families. We either reuse the submitted choice or select with surrogates matched to each family. The pool oracle uses target labels only to measure the selection gap.

Table R3. Paired protocol AUROC.

Target familyReused choiceMatched surrogatesPool oracle
Removal only60.167.268.3
Addition only60.366.066.7
Degree preserving rewiring64.070.471.8
Class mixing additions59.865.366.3

Matched selection gains 5.5 to 7.1 points over reused choice and stays within 0.7 to 1.4 points of the oracle. The same mechanism works across these families when the surrogates represent the anticipated change.

Q4. Discrepancy score

We appreciate the close look at the score. It sends the same node confidence through current hyperedge membership and a soft view built from representations and clean prototype centers. The latter does not reuse current hyperedges, so the difference traces relational disagreement for the same node evidence.

Table R4. Standalone score AUROC.

VariantPairedMixed q = 0.3
Submitted discrepancy59.160.1
Random prototypes59.560.2
kNN view instead of prototypes51.256.5
No current incidence propagation45.652.6

The contrast matters more than exact prototype placement. The discrepancy is selected in 41 of 144 runs, and removing it from the pool lowers HGNN AUROC by 1.4 points.

Q5. Recent methods and source of the gain

We thank the reviewer for requesting recent comparisons. The submission includes NODESAFE (2024), HyperGOOD (2026), and UniGOD adapted to hypergraphs (2026). We added GRASP, SCALE, NECO, and fDBD (2024), plus ASH (2023). Table R5 uses the 24 settings where all five apply. Its mixed column uses q = 0.3 and is a subset of Table R2.

Table R5. AUROC against added methods.

MethodPairedMixed
ASH57.056.0
SCALE60.359.1
NECO51.354.0
fDBD60.060.8
GRASP62.652.5
HADES, submitted selection69.669.2
HADES, mixed node surrogatesN/A76.9

We also compared selection with cross dataset choice and score averaging. Mean ranks in Table R6 are among the seven listed rules. Only the oracle uses target labels.

Table R6. Selection rules over 48 settings.

RuleAUROCMean rank
Pool oracle68.81.76
HADES self validation67.92.28
Best fixed detector on other datasets62.83.97
Best pool candidate on other datasets60.04.43
Average after z scoring60.24.77
Average after percentile scoring61.04.83
Random pool candidate55.35.96

The selector gains 5.1 points over transferring the best fixed detector from other datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.