acceptodds
Under review as a conference paper at ICLR 2027

PANDA: Learning the Hypergraph You Were Never Given via Prototype-Anchored Diffusion

Abstract

Hypergraph neural networks leverage higher order interactions by propagating node features along an observed incidence. When the observed incidence misses task relevant relations, hypergraph structure learning methods infer or refine hyperedges from node features. When learned and observed structures share a propagation schedule, their diffusion depths cannot be controlled independently. We introduce Prototype Anchored Neural Diffusion for Adaptive Hypergraphs (PANDA), which learns a soft incidence between nodes and trainable feature space prototypes. Shallow diffusion over its induced operator limits repeated global coupling, while an observed view filter with independently chosen depth supports either one step smoothing or deeper restart propagation. A learned scalar gate combines the two filtered representations at each layer. The learned incidence requires storage, and neither filter needs to materialize a dense node to node operator. Across 12 node classification benchmarks, PANDA achieves the best average accuracy rank of , compared with for the runner up TF-HNN. The results support allocating diffusion depth according to relational connectivity, so learning global relations need not entail deep global mixing.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hyergrapgh, Prototype, Anchor, Diffusion

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_9QEBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Prototype Anchored Neural Diffusion for Adaptive Hypergraphs (PANDA) for hypergraph node classification. PANDA learns a soft incidence between nodes and trainable feature space prototypes. Each layer combines two filtered representations: shallow diffusion and an observed view filter. PANDA achieves the best node classification performance in the majority of evaluation datasets.

Critical strengths

(1) The proposed method is evaluated in extensive hypergraph datasets and demonstrates the effectiveness in hypergraph node classification datasets.

(2) The mathematical formulation of each module is easy to follow.

Critical weaknesses

(1) Since the prototype is calculated on the whole node set, the proposed method is transductive. For unseen nodes, the method cannot be applied directly.

(2) The paper is evaluated on 12 datasets, while the largest dataset has 32,000 nodes. But the capability and cost (memory and inference speed) of the proposed method need to be further evaluated on huge dataset, such as OGB.

(3) The "diffusion" word looks like the spectral message passing mechanism in the previous hypegraph neural network works, instead of the image diffusion generation methods. It may be confusing to follow the core mechanism of the paper.

(4) PANDA achieves 100.0 macro-F1 on mushroom and 100.0 accuracy on zoo. And I do not find any analysis related to these results. Is it convincing?

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

We evaluated PANDA on unseen nodes and larger graphs, clarified diffusion, and corrected the zoo result.

Q1

Thank you for raising the unseen node setting. PANDA has no node specific trainable parameters. A new node receives prototype memberships from its features, and its observed memberships enter the hypergraph operator available at inference. We tested this by removing all test nodes and their incidences during training. Prototypes were initialized using training graph features only. The trained weights were frozen before test nodes and their memberships were introduced.

DatasetPANDA inductivePANDA transductiveHGNN inductiveUniGCNII inductiveMLP
cora_ca70.2 ± 1.471.7 ± 0.862.2 ± 1.169.9 ± 1.052.5 ± 1.1
cooking20037.2 ± 0.538.4 ± 0.320.4 ± 2.217.3 ± 0.318.6 ± 0.1
dblp_ca87.8 ± 0.688.9 ± 0.287.5 ± 0.386.8 ± 0.378.3 ± 0.2
modelnet4098.0 ± 0.198.0 ± 0.292.1 ± 0.297.7 ± 0.196.3 ± 0.1

These are test accuracies over five fresh training runs. The transductive column is a paired rerun using the same implementation. Across all 12 datasets, PANDA leads the four evaluated models on seven and ranks second on four. Its inductive accuracy is within two points of its transductive accuracy on ten datasets. Inference uses the features and available observed memberships of new nodes, with no new labels. The feature branch remains defined if a new node has no observed memberships.

Q2

Thank you for requesting a larger scale evaluation. The largest benchmark in the paper is dblp_ca with 41,302 nodes. The 32,000 node figure describes the controlled CPU scaling experiment. We additionally measured a graph propagation adaptation on ogbn arxiv with 169,343 nodes, and PANDA on synthetic hypergraphs with up to one million nodes.

ogbn arxiv modelTest accuracyInference timeTraining peak memory
APPNP observed operator71.06 ± 0.18186 ms1,034 MB
APPNP plus feature operator71.17 ± 0.19198 ms1,465 MB
Synthetic hypergraph nodesInference timeTraining stepTraining peak memory
300,000288 ms1,096 ms2.85 GB
1,000,0001,476 ms4,779 ms9.45 GB

The arxiv experiment uses the official split and a graph adaptation. The synthetic experiments use the hypergraph operator and were timed on an otherwise idle GPU. Memory reports peak allocated tensor memory. Full batch PANDA exceeded the 11.5 GB process cap at two million nodes. Arxiv timings were collected on a shared GPU and are indicative.

Q3

Thank you for noting the possible ambiguity. Diffusion in this paper means repeated propagation by a normalized hypergraph operator, following the terminology of personalized PageRank propagation and hypergraph diffusion. It does not denote a denoising generative model. We will define the term at its first use in Section 3.2.

Q4

Thank you for asking us to examine the near perfect results. Mushroom has 2,031 test nodes. Its PANDA accuracy is 99.9%. The previously displayed 100.0% Macro F1 is a rounded value of 99.9606%. A decision tree and a nearest neighbor classifier trained only on node features also classify every mushroom test node correctly. This dataset is highly separable under the evaluated split.

The zoo dataset only has 26 test nodes. The previous run that achieved 100.0% accuracy was due to using an earlier version. The updated results are as follows.

DatasetAccuracyMacro F1Test nodes
mushroom99.9 ± 0.199.9 ± 0.12,031
zoo93.8 ± 4.685.9 ± 7.226

Thank the authors for their response. Their response has addressed my concern to some extent. While the innovations in this paper are not particularly compelling to me, so I will maintain my original score.

Thank you for your guidance. The comments you and the other reviewers provided have greatly improved my paper. I will design and propose more innovative algorithms in the future.

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

The paper addresses learning in hypergraphs. It aims to unify two scenarios, one when the hypergraph structure is given (observed), and one, when it needs to be inferred from the node features. It claims that there are problems that gain from applying these two views in parallel, but that they should be clearly separated being favored by different GNN diffusion patterns. The work extends the state of the at this point. Otherwise it builds on earlier works on learning over observed hypergraphs and inference of hypergraph structures.

The proposed structure with the parallel branches is motivated with a somewhat intuitive reasoning by comparing the support the observed and inferred branch have.

The proposed method, PANDA is evaluated extensively against several methods of several classes of approaches that could be used, and it shows that PANDA is among the best performing ones for all the baselines. In addition the authors perform several interesting evaluations with the aim to understand how inferring hypergraph structure works, how the hyperparameters should be set and performs a rigorous ablation study.

Critical strengths

Since the proposed PANDA outperforms baselines in node classification accuracy in most of the cases, this is definitely a strength of the proposal. The gain is mainly a couple of percentage improvement over the best baseline, but this best baseline depends on the dataset. So all in all, the proposed method seems to perform well independently from the characteristics of the dataset.

The paper also performs thorough evaluation, covering most of the aspect one could think of to understand how inferring hypergraph structure improves the learning performance. I appreciate especially the ablation study, and the question on what the prototypes represent.

Critical weaknesses

Even though PANDA leads to higher or comparable accuracy than the competitors, it is not a transformative technique. If the accuracy results are low for the baselines for a given dataset, then it remains in the same low region also for PANDA.

A weakness that could be addressed in a revision is the focus of the presentation and the evaluation. As I see it, the main contribution of the paper is the two parallel learning branches, one for the observed and one for the inferred hypergraph, and the different diffusion structure the authors propose. This proposal should be discussed more deeply, now there is only a somewhat intuitive discussion of the supports and then an evaluation of the hyperparameters at the end of the appendix, in Appendix D.5.3. Considering the parallel branches, maybe I interpret Figure 15 incorrectly, but it seems to me that "observed" would gain applying m>2 in several cases. Once higher m is selected, then the simple structure of the diffusion in (6) is maybe not motivated.

Questions for authors

I am wondering what the improvement effect of feature-induced hypergraphs actually means. This paper is set in hypergraph context. But does this technique need to be combined with observed hypergraphs or could (or should) be added to any graph learning framework? This could be clarified in the paper.

The observed view part of Figure 2 is confusing. The coloring does not seem to be consequent, and as I understand, the last column of the augmented observed view should contain on one colored square only.

The reasoning about the dual diffusion in Section 3.2 should be discussed again based on the results of Appendix D.5.3.

I find the ablation study interesting, but while the numbers are presented, they are not really discussed. Would it be possible to discuss these results more extensively? For example, does the parameter M of the datasets (Table 4) correlated to the importance of the building blocks of PANDA?

The same holds for Appendix D.5.3. These results should support the main idea of the paper. But do they? More reflections are needed.

3 comments
AuthorsOpenReview ↗

We clarify the role and transferability of the feature induced view, correct Figure 2, and add ablation and diffusion results to explain PANDA’s design.

1. Feature induced hypergraphs

Thank you for asking about transfer beyond hypergraphs.

The learned N×KN\times K incidence defines soft hyperedges and can propagate without observed memberships. Table R1 reports the six main ablation datasets using the full model values in the paper. All entries are five seed accuracy means in percent.

DatasetFeature onlyObserved onlyBoth
cora_ca59.471.273.0
citeseer_cc58.158.660.8
cora_cc59.468.369.1
ntu201286.887.590.5
modelnet4096.192.798.0
house53.655.056.2

We also added the feature view to GCN and APPNP on eight ordinary graphs under one shared configuration. Table R2 gives all eight results.

GraphGCNGCN plus featureAPPNPAPPNP plus feature
cora80.179.279.379.9
citeseer64.765.464.865.9
pubmed75.876.577.478.8
chameleon66.059.753.849.8
squirrel52.240.538.534.3
actor29.329.530.933.0
roman-empire47.170.040.372.1
amazon-ratings49.549.147.948.8

Across all eight graphs, the added view improves four GCN and six APPNP results. It can extend graph propagation, while the observed hypergraph remains valuable for PANDA's target tasks.

2. Figure 2

We appreciate the careful inspection of Figure 2.

We corrected the observed view drawing. Each appended singleton column contains exactly one colored square in its corresponding node row. The hypergraph, row colors, node degrees, and augmented incidence now agree, including the isolated fifth node.

3. Dual diffusion in Section 3.2

Thank you for revisiting the diffusion rationale.

Appendix D.5.3 applies the saved operators repeatedly to centered random probes. Table R3 gives the fraction of feature variation left after one application and the ratio of observed variation at steps ten and one.

DatasetFeature after step 1Observed step 10 / step 1
cora_ca0.0120.71
citeseer_cc0.0170.91
pubmed_cc2.7×10−82.7\times10^{-8}0.98
modelnet400.0520.25

The feature operator supplies global support in one step and can strongly contract centered variation. The observed operator expands paths only within the supplied components. Section 3.2 therefore gives the two operators independent finite filters. The diagnostic motivates separate control, while predictive comparisons below assess the chosen schedules.

4. Interpreting the ablation

We appreciate the question about dataset structure.

The hyperedge count MM does not by itself predict component importance. Table R4 gives approximate Spearman associations from the displayed ablation means aligned with the main paper. Singleton and observed depth analyses exclude cooking200 because its selected setting uses neither component. Costs are accuracy decreases in points.

Removed componentDatasetsAssociation of MM with cost
Feature view12−0.06-0.06
Observed view12+0.28+0.28
Singleton augmentation11−0.23-0.23
Fusion gate12−0.46-0.46
Selected multistep observed branch11+0.06+0.06

Membership structure is more informative in specific cases. Citeseer_cc has 56 percent isolated test nodes and loses 4.3 points without singleton edges. Modelnet40 has none and loses 0.6 point. Modelnet40 gains 5.3 points from the feature view despite having 10,630 hyperedges, whereas cooking200 gains essentially none with 2,750. These cross dataset associations are descriptive.

5. What Appendix D.5.3 supports

Thank you for asking what D.5.3 establishes.

Table R5 tests predictive accuracy under separate filters, two shared depths, and operator mixing before diffusion. Separate uses m=2m=2 and L=10L=10, except cooking200 with L=1L=1. Mix first applies the observed filter to a weighted operator combination. All other settings stay fixed. Values are five seed means from controlled reruns, separate from the main results.

DatasetSeparateEqual 2Equal 10Mix first
cora_ca72.570.872.772.6
citeseer_cc60.761.360.660.6
cora_cc68.966.269.571.2
cooking20037.928.329.737.9
zoo93.8100.087.787.7
house54.751.655.051.9
ntu201290.089.889.088.7
modelnet4098.198.397.395.7

Cooking200 favors its selected one step observed filter over either equal depth restart alternative. With observed depth held at 10, increasing feature depth to 10 lowers zoo by 6.1 points, ntu2012 by 1.0, and modelnet40 by 0.8. Mixing before diffusion lowers modelnet40 by 2.4 and ntu2012 by 1.3. Other datasets favor different choices, so the results support independent control rather than one shared schedule for every task.

I appreciate the additional numerical resutls the authors provided, it answered some of my questions. I still belive though that a more extensive evaluation of diffusion depth would benfit the paper. It is also unfortunate that the results are not consistent across the datasets, it seems to be hard to establish when PANDA can lead to significant gains.

AuthorsOpenReview ↗

We evaluated both diffusion depths across all 12 datasets and used controlled experiments to identify when PANDA gains from combining feature and observed relations.

Thank you for clarifying both concerns. We expanded the depth study to all 12 datasets and tested when the two relation sources yield a substantial gain.

We evaluated all 25 combinations of feature induced depth m∈{1,2,3,5,10}m\in\{1,2,3,5,10\} and observed depth L∈{1,2,5,10,20}L\in\{1,2,5,10,20\} with five fresh models per cell. This comprises 1,500 runs. All other settings were held fixed within each dataset. Table R6 reports accuracy ranges and changes in percentage points. All contrasts are calculated within this controlled run series. Its reference schedule uses (m,L)=(2,10)(m,L)=(2,10) except on cooking200, which uses (2,1)(2,1). The validation choice maximizes mean validation accuracy over the 25 cells.

DatasetRange over mmRange over LLValidation choiceTest change
cora_ca0.61.5(10, 5)+1.9
citeseer_cc0.81.5(3, 10)0.0
cooking2000.814.0(5, 1)+0.3
cora_cc0.73.9(2, 10)0.0
pubmed_cc0.20.1(5, 2)+0.1
dblp_ca0.21.5(2, 5)+0.2
zoo3.810.8(10, 1)+6.2
20news0.20.5(1, 2)+0.4
mushroom0.10.2(2, 10)0.0
ntu20120.52.4(5, 2)+0.9
modelnet400.12.3(1, 5)+0.1
house1.62.8(2, 10)0.0

Feature induced depth changes accuracy by at most 0.8 points on ten datasets. The observed branch has a larger and less uniform response. Cooking200 obtains 38.2 at L=1L=1, while every tested L≥2L\geq2 obtains at most 30.3. Cora_cc rises from 66.8 at L=1L=1 to 69.2 at L=10L=10, and modelnet40 rises from 96.1 to 98.0. The L=1L=1 branch uses one step HGNN smoothing, whereas L≥2L\geq2 uses restart diffusion, so this comparison measures the observed filter choice as well as depth. Selecting both depths by validation changes test accuracy by at most 0.4 points on nine datasets. Zoo is the largest exception. It has 26 test nodes, so one error changes accuracy by 3.85 points.

To examine variation across datasets, we independently varied feature signal, hyperedge informativeness, and the share of nodes eligible for observed hyperedges in synthetic hypergraphs with 2,000 nodes. PANDA, HGNN, and an MLP used fixed settings across all conditions. Table R7 reports PANDA's five seed mean accuracy margin over the better of HGNN and MLP, in points. Each cell also shows the number of seeds out of five on which PANDA beat both.

Eligible nodesFeature signal μ\muRandom hyperedges h=0h=0h=0.4h=0.4h=0.6h=0.6h=0.8h=0.8
50%1-1.7 (0)+2.9 (5)+7.2 (5)+10.0 (5)
50%2-3.9 (1)+0.7 (3)+8.9 (5)+16.6 (5)
100%2-7.4 (0)-1.5 (2)+9.5 (5)+6.4 (5)

The largest gains appear when features and observed memberships both contain class information. Partial observed coverage can amplify the gain. At 50% eligible nodes and μ=2\mu=2, raising hyperedge informativeness from h=0h=0 to h=0.8h=0.8 changes PANDA's margin from -3.9 to +16.6 points. At h=0h=0, PANDA trails the stronger simple model. At h=0.6h=0.6 and h=0.8h=0.8, it exceeds both on every seed in these three rows. This interaction gives a concrete interpretation of the uneven ablation gains across datasets.

Our correction to the earlier zoo result remains 93.8±4.693.8\pm4.6 accuracy and 85.9±7.285.9\pm7.2 Macro F1, as reported to Reviewer 3. That result and the depth grid are separate fresh model runs. We will synchronize the main and appendix tables in the revision and label the controlled depth study separately.

Reviewer_N7viOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents PANDA, a transductive hypergraph node-classification architecture that separates feature-induced relations from the observed incidence. A trainable encoder and K prototypes define a row-normalized N-by-K soft incidence, inducing a positive-semidefinite, rank-constrained operator with global support. PANDA applies a shallow polynomial filter to this operator, while independently applying either one-step smoothing or truncated restart propagation to the observed hypergraph. A learned scalar gate fuses the two outputs at each layer. The factorization avoids materializing an N-by-N operator, and optional singleton hyperedges preserve isolated nodes. Experiments on 12 benchmarks compare against 22 baselines. PANDA reports an average accuracy rank of 2.00 and an average Macro-F1 rank of 2.67, supported by component ablations, corruption studies, and efficiency measurements.

Critical strengths
  1. The asymmetric treatment of the two relation sources is well motivated and convincingly analyzed. Repeated propagation has different meanings for a globally supported prototype operator and an observed hypergraph operator. The paper turns this observation into a coherent design: the learned branch uses shallow diffusion, the observed branch has an independently selected schedule, and fusion happens only after filtering. The spectral analysis and the singleton construction for isolated nodes further strengthen the presentation.

  2. The empirical evaluation is broad and thorough. Beyond the 12-dataset comparison, the paper includes Macro-F1 results, cross-dataset statistical tests, component removal on 11 datasets, sensitivity analyses, hyperedge removal/rewiring, and controlled scaling measurements. The ablations support the contribution of both views and clarify when singleton augmentation matters. The efficiency results also directly support the practical motivation for the factorized incidence.

Critical weaknesses
  1. The comparison with the closest structure-learning methods is incomplete and not always like-for-like. Several methods discussed as closely related, including SPHINX, DSA-HG, HNDiffN, and HyperAim, are not evaluated. HERALD and SHSL also fall back to observed-incidence HGNN propagation on some large datasets, while other methods have missing entries. The paper is transparent about these adaptations, and the complete-method statistical analysis is useful. Nevertheless, the current results do not fully establish the advantage over the most relevant recent alternatives under comparable tuning budgets.

  2. The conceptual novelty is more limited than the empirical package. Prototype/anchor factorization, restart-based propagation, and gated dual-branch fusion are established components. The contribution lies primarily in combining them around the distinction between repeated global coupling and reach along observed memberships. I find this distinction useful, but the paper would benefit from positioning the algorithmic advance more precisely against anchor-based structure learning and independently filtered multi-branch architectures.

Questions for authors
  1. Can you add the closest recent structure-learning baselines, or explain their exclusion? For the adapted HERALD/SHSL entries, how were hyperparameters selected, and can you provide a comparison in which all methods receive comparable dataset-specific validation search?

  2. The main experiments use one data split per dataset with five optimization seeds. Do the main accuracy and Macro-F1 conclusions persist across independently generated train/validation/test splits, especially for the small datasets and those with many isolated test nodes?

3 comments
AuthorsOpenReview ↗

We add SPHINX and DSA-HG comparisons, clarify code availability and baseline settings, correct the SHSL results, and evaluate performance across ten independent data splits.

Response to Reviewer 1

Q1. Recent baselines and hyperparameters

Of the four suggested methods, only SPHINX has publicly released code. We compare SPHINX with a paper based reproduction of DSA-HG, another structure learning method. HNDiffN and HyperAim have no public implementations. PANDA uses its original validation selected configurations, without retuning against these comparisons. Table R1 summarizes the displayed dataset means. W/T/L counts PANDA wins, ties, and losses. Table R2 gives the five seed accuracy means beside PANDA and TF-HNN, the prior runner up.

Table R1. Coverage and dataset outcomes.

MethodImplementationCompletedAccuracy W/T/LMacro-F1 W/T/L
SPHINXReleased code12 / 129 / 1 / 210 / 0 / 2
DSA-HGPaper based implementation12 / 1212 / 0 / 010 / 0 / 2

Table R2. Test accuracy in percent.

DatasetPANDATF-HNNSHSLSPHINXDSA-HG
cora_ca73.070.072.354.860.8
citeseer_cc60.859.056.952.231.8
cooking20038.017.119.218.620.3
cora_cc69.165.067.854.851.4
pubmed_cc39.339.233.639.137.9
dblp_ca88.985.688.478.087.2
zoo100.086.979.291.550.0
20news81.882.482.282.381.7
mushroom99.999.899.799.998.6
ntu201290.590.578.187.284.3
modelnet4098.098.291.096.291.4
house56.255.454.557.549.8

DSA-HG omits its unspecified incremental update.

We replaced the earlier adapted SHSL row with its released node classification path. It trains UniSAGE on the anchor hypergraph without passing learned hyperedges or contrastive embeddings to the classifier. PANDA exceeds the corrected SHSL result on 11 of 12 datasets. Table R3 specifies the configurations of the two methods raised by the reviewer.

Table R3. Baseline settings and memory adaptations.

MethodTraining configurationImplementation detail
HERALD128 hidden units, learning rate 0.001, weight decay 0.0005, dropout 0.5, blending 0.1Observed incidence HGNN fallback on pubmed_cc, dblp_ca, 20news due to memory
SHSLReleased UniSAGE path, 32 hidden units, learning rate 0.01, no weight decay, dropout 0.5Corrected released path on all 12 datasets

A comparable validation search would use a method specific search space with the same trial budget for each method. We did not conduct that additional search. Our comparisons share splits, seeds, and checkpoint selection, while using available method configurations and stated backbone settings otherwise. PANDA was also run with one shared configuration across all 12 datasets.

Q2. Independent data splits

We tested ten independently generated splits on zoo, house, ntu2012, cora_ca, cora_cc, citeseer_cc, and pubmed_cc. The four citation datasets retain their per class training budgets and resample 500 validation nodes. The other three use 50/25/25 splits. Each split trains a fresh model. PANDA keeps its original configuration across splits, so this also tests transfer without new parameter selection. Nine baselines and a shared PANDA configuration use the same splits. Table R4 reports mean rank across the ten splits and seven datasets, with lower rank better.

Table R4. Mean rank across independent splits.

MetricPANDAPANDA shared settingBest other method
Accuracy3.033.16TF-HNN 3.30
Macro-F12.933.13UniGCNII 3.63

Across this targeted set of small and highly isolated benchmarks, PANDA retains the best aggregate rank for both metrics. The shared setting also ranks ahead of every baseline. The conclusion concerns aggregate ranking. citeseer_cc remains an exception, with DHGNN stronger on isolated nodes.

Thank you for the detailed response and the additional experiments. The results across ten independent splits largely address my robustness concern, and the added SPHINX/DSA-HG comparisons and corrected SHSL results improve the baseline coverage. However, a comparable dataset-specific validation search across methods was not conducted, so this aspect of my concern about fair comparison remains only partially addressed. I encourage the authors to include these new results and baseline details in the revision, clarify the remaining tuning-budget limitation, and further sharpen the discussion of novelty relative to prior anchor-based and multi-branch methods.

AuthorsOpenReview ↗

We added a matched validation search across five methods, report the results and remaining tuning limits, and clarify PANDA’s distinction from prior anchor based and multiple branch methods.

Thank you for the follow up and for recognizing the new split and baseline experiments. We have now run a matched validation search to address the remaining comparison concern.

We attempted the same 15 candidates for PANDA, TF-HNN, UniGCNII, SheafHyperGNN, and HyperGCN on every dataset. The grid combines hidden widths 64, 128, and 256 with learning rates 0.001, 0.005, 0.01, 0.02, and 0.05. Every candidate used five fresh models, the paper's split and training protocol, and the method's default settings for all other parameters. We selected the highest mean validation accuracy, with validation Macro F1 breaking ties. Test results did not enter selection. PANDA used one shared base configuration for all datasets, with its prototype count, propagation depths, and singleton policy fixed.

Five seed mean test accuracy in percent after validation selection is shown below. The final column compares average ranks before and after search among these five methods on their available datasets.

Methodcora_caciteseer_cccooking200cora_ccpubmed_ccdblp_ca
PANDA72.561.030.469.339.387.0
TF-HNN71.459.919.366.439.285.6
UniGCNII73.160.919.069.238.386.9
SheafHyperGNN70.959.618.365.739.4OOM
HyperGCN59.059.818.564.439.180.5
Methodzoo20newsmushroomntu2012modelnet40houseAverage rank, default to searched
PANDA98.582.3100.090.598.356.61.75 to 1.42
TF-HNN93.882.699.990.498.356.42.42 to 2.42
UniGCNII90.082.599.890.298.054.52.75 to 2.83
SheafHyperGNN90.0OOMOOM89.1OOM57.23.25 to 3.50, eight datasets
HyperGCN96.282.699.985.996.356.23.17 to 3.58

PANDA ranks first or tied first on eight datasets and no lower than third on any. The matched search raises mean test accuracy by 0.90 to 1.58 points across the five methods. PANDA has the lowest average rank among the four methods that complete all 12 datasets. HyperGCN now rebuilds its stochastic projection graph for each seed, which corrects the reuse in our earlier evaluation.

The remaining tuning budget limit is that this grid varies only width and learning rate. Method specific depth, dropout, regularization, and architectural choices stay at their defaults. SheafHyperGNN exceeds memory on four datasets, and only five of its candidates fit on cooking200.

We also agree that the relation to anchor based and multiple branch methods should be sharper. AnchorGraph and IDGL already use anchor assignments, while DSA-HG induces latent hyperedges. PANDA's contribution is the order of computation. It learns a feature induced incidence while retaining the observed hypergraph as a separate operator, applies a shallow filter to the globally supported learned operator and an independently chosen restart filter to the observed operator, then gates their outputs in each layer. IDGL, HERALD, and DSA-HG combine learned and observed relations before propagation. S2-HGNN constructs its branches from expansions of the observed hypergraph. PANDA's learned branch can relate nodes outside the observed incidence.

Controlled schedule comparisons show the effect of this separation. A common depth of two loses 2.7 points on cora_cc, while a common depth of ten loses 1.0 point on ntu2012. Combining the operators before diffusion loses 2.3 points on modelnet40 and 1.3 points on ntu2012, but improves cora_cc by 2.3 points. This variation supports giving the two operators separate schedules. We will add the search and its full configurations, the ten split results, the related method comparison, and the SPHINX, DSA-HG, and corrected SHSL details to the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.