acceptodds
Under review as a conference paper at ICLR 2027

Hypergraph Structural Entropy Network for Attributed Hypergraph Clustering

Abstract

Attributed hypergraph clustering leverages high-order relationship modeling to deliver discriminative vertex representations for data grouping. However, precise clustering is often hindered by the decoupling of the representation learning and grouping stages. Moreover, prevailing contrastive learning methods rely heavily on meticulously designed augmentations to achieve invariance, without explicitly guaranteeing a cluster-separable embedding space. To address these limitations, we propose the Hypergraph Structural Entropy Network (HSENet), a unified generative clustering framework that directly couples representation learning with precise data grouping. At its core, we formulate hypergraph structural entropy (HSE) to explicitly capture the hierarchical topology of hypergraphs, generalized for edge-dependent vertex weights (EDVW) and equipped with a continuous, almost-everywhere differentiable relaxation via the Lovász extension. Within a joint, end-to-end framework, HSENet couples two objectives: a masked generative self-supervised module to learn vertex semantics and structural topology, and an HSE-guided clustering assigner that optimizes a soft partitioning tree on an enhanced representation hypergraph. Extensive experiments on 8 benchmark datasets demonstrate that HSENet outperforms 10 competing methods, highlighting its clustering superiority.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Attributed Hypergraph Clustering, Hypergraph Structural Entropy, Structural Information Theory

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_epBFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes HSENet for attributed hypergraph clustering. It formulates hypergraph structural entropy for hypergraphs with edge-dependent vertex weights and derives a continuous Lovász-extension relaxation for learning soft hierarchical assignments. The framework jointly trains masked feature reconstruction, hyperedge filling, representation-based topology enhancement, and an entropy-guided clustering module. Experiments compare HSENet with ten baselines on eight datasets. HSENet achieves the best reported mean performance across all four metrics on five datasets.

Critical strengths

S1: The proposed objective explicitly evaluates how entire hyperedges are partitioned while accommodating edge-dependent vertex weights. Its Lovász relaxation agrees with the discrete splitting costs at binary assignments and enables gradient-based optimization. This provides a concrete mechanism for connecting representation learning to hypergraph partition quality.

S2: In the empirical evaluation, improvements across multiple datasets and metrics, together with component ablations and missing-attribute experiments, provide evidence that the proposed method is effective.

Critical weaknesses

W1: The experiments do not compare existing methods that explicitly optimize graph partitions. The paper discusses LSEnet, DeSE, ASIL, and SCISE but evaluates none of them. These methods could be considered for evaluation through a hypergraph-to-graph conversion. Although conversion can change the partitioning objective, that difference makes the comparison informative for assessing the benefit of directly modeling hyperedge splitting costs. The evaluation does not compare against methods that directly optimize structural objectives, such as MinCutPool (Bianchi et al., 2020) and DMoN (Tsitsulin et al., 2023). Evaluating these methods on a graph expansion of the same hypergraphs would help establish the benefit of the proposed hyperedge-based entropy objective.

Bianchi, F. M., Grattarola, D., and Alippi, C. (2020). “Spectral Clustering with Graph Neural Networks for Graph Pooling.” ICML.

Tsitsulin, A., Palowitch, J., Perozzi, B., and Müller, E. (2023). “Graph Clustering with Graph Neural Networks.” JMLR, 24(127), 1–21.

W2: The ablations do not sufficiently isolate the contribution of HSE. The k-means variant removes the hierarchical assigner, HSE loss, and orthogonality loss together, and retrains the encoder using only the generative objectives (Appendix E.1). Consequently, this comparison changes several components simultaneously. Improvements over this variant are uneven: Table 2 reports relatively small gains on Cora and higher NMI for k-means on 20News, although larger gains occur on AMiner. A controlled comparison retaining the encoder, enhanced topology, and assignment architecture while replacing HSE with another clustering objective would more clearly establish the value of the principal contribution. The current results support the effectiveness of the overall system more strongly than the specific advantage of its entropy objective.

Questions for authors

None.

7 comments

W1: Comparison with graph partitioning methods

Thank you for this suggestion. The added experiments show that HSENet achieves the highest mean NMI and ACC among methods with available results on all 8 datasets.

We added a new subsection, “Structural Objectives and Graph Clustering Baselines,” where we evaluate all 6 requested methods: MinCutPool, DMoN, LSEnet, DeSE, ASIL, and SCISE. The graph baselines operate on binary clique expansions of the same hypergraphs, while HSENet operates directly on hypergraphs. This comparison includes both cut-/modularity-based methods and graph structural entropy methods.

Table R1. Graph clustering comparison. Results (%) are mean ± standard deviation over five runs.

MetricMethodCoraCiteseerPubmedDBLP20NewsMushroomAMinerZoo
NMIMinCutPool27.8±14.020.5±2.420.6±1.153.2±2.4OOM4.5±9.033.0±1.00.0±0.0
NMIDMoN26.8±13.818.9±2.40.2±0.238.7±8.4OOM0.0±0.038.8±5.20.0±0.0
NMISCISE53.1±1.139.9±0.633.6±0.559.4±2.929.7±1.521.4±0.442.9±0.372.8±1.9
NMIDeSE39.8±3.722.1±2.317.3±8.3OOMOOMOOMOOM0.0±0.0
NMILSEnet29.4±7.324.4±5.525.1±7.5OOMOOMOOMOOM14.1±0.6
NMIASIL32.0±3.822.4±1.924.4±2.5OOMOOMOOM45.9±4.910.9±0.7
NMIHSENet57.5±0.845.6±0.635.2±2.065.8±2.643.5±1.245.8±0.149.9±0.990.5±1.9
ACCMinCutPool45.4±9.644.4±2.458.5±5.966.5±4.9OOM56.9±10.129.4±1.340.6±0.0
ACCDMoN42.8±8.641.4±2.441.9±1.552.9±6.9OOM51.8±0.036.0±3.540.6±0.0
ACCSCISE71.7±0.864.1±0.972.8±0.472.3±6.161.6±1.275.9±0.235.7±0.673.3±3.3
ACCDeSE58.9±6.847.7±4.057.9±8.9OOMOOMOOMOOM40.6±0.0
ACCLSEnet41.0±11.246.7±7.863.7±8.6OOMOOMOOMOOM45.0±0.8
ACCASIL45.6±2.645.9±3.555.6±0.6OOMOOMOOM41.1±3.731.5±0.5
ACCHSENet74.7±1.672.1±0.474.0±2.078.6±5.576.0±1.086.8±0.043.1±0.693.1±0.6

HSENet achieves the highest NMI and ACC among methods with available results on each dataset, with NMI gains of 24.4 and 13.8 percentage points over the strongest available graph baseline on Mushroom and 20News, respectively. On these 2 datasets, expansion produces dense graphs (21.1% and 20.2% density), and DeSE, LSEnet, and ASIL encounter OOM on both. Zoo becomes a complete graph, whose cut costs depend only on cluster sizes and cannot distinguish equally sized vertex groups; this loss of discriminative topology may contribute to the low NMI of several baselines. These comparisons support HSENet's overall clustering advantage over clique-expansion pipelines. More results are added in Section 6.3 and Appendix E.2.

W2: Controlled evaluation of the HSE objective

The controlled objective replacements show that HSE achieves the highest average scores on all 4 metrics and the best average rank under the fixed HSENet training configurations.

We agree that the original kk-means variant changes several components simultaneously and therefore evaluates the combined assignment-learning design. We have added the requested comparison: retaining the encoder, hierarchical assignment architecture, topology-enhancement procedure, reconstruction and hyperedge-filling losses, and orthogonality regularizer, we replace only the structural objective with differentiable hypergraph normalized cut (NeurIPS 07 & 17), strict hypergraph modularity (Science Advances 21), and Graph SE. For Graph SE, a weighted clique expansion of the enhanced EDVW hypergraph is used solely to evaluate the loss.

Table R2. Controlled structural-objective comparison. Scores (%) are equally averaged over 8 datasets.

Structural objectiveNMIARIACCMacro-F1Avg. rank
Hypergraph NCut51.6450.4371.7464.792.50
Hypergraph Modularity48.6245.0465.8751.213.44
Graph SE52.1751.2572.2164.572.50
HSE54.2154.2974.8067.781.56

HSE achieves the highest average scores across the 8 datasets on all 4 metrics, with particularly substantial gains over hypergraph modularity, ranging from 5.59 to 16.58 percentage points. HSE also achieves the best average rank of 1.56, whereas Graph SE and hypergraph NCut tie at 2.50. HSE's gains over Graph SE may reflect the benefits of directly modeling hyperedge splitting costs. Together, these results support HSE as an effective structural objective and indicate that it is an important contributor to HSENet's overall clustering performance. This comparison is added in Section 6.3. Objective definitions and detailed results are added in Appendix D.4 and E.1.

AuthorsOpenReview ↗

Kind Reminder: Rebuttal and Reviewer Feedback

Dear Reviewer,

Thank you for your valuable comments. As the discussion deadline is approaching, we would greatly appreciate it if you could revisit our rebuttal. We have added the required comparisons, in which HSENet and the HSE objective consistently achieve the highest performance. We hope these results address your concerns.

Please let us know if you have any remaining concerns. Thank you again for your time and consideration.

AuthorsOpenReview ↗

W1: Detailed comparison with graph partitioning methods

Thank you again for this helpful suggestion. Following our brief response, which referred to the updated manuscript PDF, we provide a more detailed response to W1 below, including the complete results added in Section 6.3 and Appendix E.1.

We agree that graph partitioning methods provide an informative comparison despite the change in objective induced by conversion. We evaluated all six requested methods: MinCutPool and DMoN, which optimize cut- and modularity-based objectives, and the graph structural entropy methods LSEnet, DeSE, ASIL, and SCISE. All graph baselines use binary clique expansions of the same eight hypergraphs; HSENet operates directly on hypergraphs.

Table R1. Complete results (%), mean±standard deviation over five runs. Bold indicates the highest mean in each row. F1 denotes Macro-F1; OOM denotes out of memory.

DatasetMetricMinCutPoolDMoNSCISEDeSELSEnetASILHSENet
CoraNMI27.8±14.026.8±13.853.1±1.139.8±3.729.4±7.332.0±3.857.5±0.8
ARI19.8±10.219.8±10.648.9±1.134.4±5.021.7±9.424.4±3.352.9±0.9
ACC45.4±9.642.8±8.671.7±0.858.9±6.841.0±11.245.6±2.674.7±1.6
F140.7±17.829.2±12.570.1±0.946.6±10.821.3±13.627.3±3.369.5±4.0
CiteseerNMI20.5±2.418.9±2.439.9±0.622.1±2.324.4±5.522.4±1.945.6±0.6
ARI17.3±2.415.9±1.439.2±0.718.2±2.521.8±6.420.9±1.847.9±0.7
ACC44.4±2.441.4±2.464.1±0.947.7±4.046.7±7.845.9±3.572.1±0.4
F142.6±2.528.8±4.659.6±1.040.6±3.834.2±11.131.1±5.161.7±0.4
PubmedNMI20.6±1.10.2±0.233.6±0.517.3±8.325.1±7.524.4±2.535.2±2.0
ARI20.8±2.30.2±0.338.4±1.015.8±7.925.5±9.822.9±3.640.1±3.0
ACC58.5±5.941.9±1.572.8±0.457.9±8.963.7±8.655.6±0.674.0±2.0
F157.8±5.424.5±6.470.9±0.341.8±8.551.6±11.145.4±1.672.1±1.8
DBLPNMI53.2±2.438.7±8.459.4±2.9OOMOOMOOM65.8±2.6
ARI45.6±3.231.3±8.456.3±4.5OOMOOMOOM61.8±5.1
ACC66.5±4.952.9±6.972.3±6.1OOMOOMOOM78.6±5.5
F160.8±13.237.7±8.469.9±6.5OOMOOMOOM78.5±4.9
20NewsNMIOOMOOM29.7±1.5OOMOOMOOM43.5±1.2
ARIOOMOOM30.8±2.0OOMOOMOOM50.3±1.2
ACCOOMOOM61.6±1.2OOMOOMOOM76.0±1.0
F1OOMOOM53.9±1.6OOMOOMOOM73.0±1.7
MushroomNMI4.5±9.00.0±0.021.4±0.4OOMOOMOOM45.8±0.1
ARI5.9±11.80.0±0.026.8±0.4OOMOOMOOM54.2±0.1
ACC56.9±10.151.8±0.075.9±0.2OOMOOMOOM86.8±0.0
F142.7±17.234.1±0.075.8±0.2OOMOOMOOM86.7±0.0
AMinerNMI33.0±1.038.8±5.242.9±0.3OOMOOM45.9±4.949.9±0.9
ARI17.1±0.827.6±3.628.1±0.7OOMOOM31.2±4.635.4±0.5
ACC29.4±1.336.0±3.535.7±0.6OOMOOM41.1±3.743.1±0.6
F127.4±1.323.7±1.027.3±0.5OOMOOM17.7±2.722.1±1.4
ZooNMI0.0±0.00.0±0.072.8±1.90.0±0.014.1±0.610.9±0.790.5±1.9
ARI0.0±0.00.0±0.064.1±4.10.0±0.04.1±0.32.3±0.591.8±3.2
ACC40.6±0.040.6±0.073.3±3.340.6±0.045.0±0.831.5±0.593.1±0.6
F18.3±0.08.3±0.063.0±2.38.3±0.022.4±2.418.9±0.278.7±1.8

HSENet achieves the highest mean NMI, ARI, and ACC among methods with available results on each of the eight datasets, and the highest Macro-F1 on six. The exceptions are Cora (SCISE: 70.1 vs. 69.5) and AMiner (MinCutPool: 27.4 and SCISE: 27.3 vs. 22.1). On Mushroom and 20News, its gains over the strongest available graph baseline are 24.4 and 13.8 percentage points in NMI, and 27.4 and 19.5 points in ARI, respectively.

Clique expansion yields dense graphs on Mushroom (21.1%) and 20News (20.2%). Zoo becomes a complete graph, whose cut costs cannot distinguish vertex groups of equal size; this loss of discriminative topology may contribute to the low NMI of several baselines. SCISE completes all eight datasets; MinCutPool and DMoN complete seven. These outcomes characterize the memory limitations of the graph clustering methods on the evaluated hypergraph datasets.

Overall, these results support HSENet's overall clustering advantage over the evaluated clique-expansion pipelines and are consistent with the benefit of directly modeling hypergraph structure. We hope these additions address the requested comparison with methods that explicitly optimize graph partitions.

Reviewer_8JXUOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a structural entropy measure for hierarchical hypergraph partitioning and discuss its connection to hypergraph conductance. Based on soft cluster assignments, the model trains on the entropy's continuous extension as part of the loss function, in addition to a hyperedge-filling task with masked features. EDVW values are computed from node representations to enhance the hypergraph structure. The resulting clustering performance is strong on eight datasets.

Critical strengths
  1. The hypergraph structural entropy computed from a discrete partitioning tree is used to train a differentiable representation learning model.
  2. The EDVW assignment is a hypergraph-specific design that contributes to the method's performance.
Critical weaknesses
  1. Evaluation results for certain baselines (AHCKA, AHRC, GRAC) are inconsistent with those reported in the original papers on the same dataset, such as DBLP. Please describe the evaluation protocol for the baseline methods, such as the choice of parameters.
  2. According to Appendix E.1, the k-means variant removes HSE loss from HSENet, reducing it to MFR loss + HF loss (basically equivalent to HypeBoy). But the performance of the k-means variant in Table 2 is still much higher than HypeBoy in Table 1. What's the difference here? Clarifying this could better solidify the contribution of HSE/HSENet over other designs from existing work.
Questions for authors
  1. How are the default algorithm parameters for HSENet selected? The learning rate is set differently for certain datasets, so its sensitivity and tuning could be explained further.
  2. The computation of HSE might be quite costly. What is the computational complexity of each epoch for HSENet?
3 comments

W1: Evaluation protocal for baselines.

General Protocol: We use official implementations and default settings for clustering baselines. For SSL baselines, we use published dataset-specific settings and evaluate embeddings with k-means following CAHC. Details are in Appendix D.2.

AHCKA, AHRC, GRAC Protocol:

We evaluate each method using the metric implementations in its official code. Our settings are:

  • AHCKA: α=0.2,β=0.5,K=10,ϵQ=0.005,Ta=200,Ti=25,τ=5\alpha=0.2,\beta=0.5,K=10,\epsilon_Q=0.005,T_a=200,T_i=25,\tau=5.
  • AHRC: α=0.2,γ=2,τ=3\alpha=0.2,\gamma=2,\tau=3.
  • GRAC: α=0.5\alpha=0.5, tol=0.1tol=0.1 (DBLP: 0.5; Cora: 0.12).

AHCKA's paper uses Ta=1000T_a=1000. Possible code/parameter updates and dataset-version differences may explain the discrepancies; Cora and Citeseer have different node counts across studies.

Below are original-paper scores (%) and HSENet means. Dataset sizes differ across papers (Cora). HSENet leads 18/20 rows numerically; AHCKA leads DBLP ARI/ACC. Thus, HSENet remains strongest overall in this descriptive comparison.

DatasetMetricAHCKAAHRCGRACHSENet
CoraNMI41.246.034.6557.5
CoraARI33.841.0—52.9
CoraACC59.269.055.3974.7
CoraF152.051.050.5069.5
CiteseerNMI39.236.032.8345.6
CiteseerARI39.736.0—47.9
CiteseerACC66.267.061.2372.1
CiteseerF161.546.057.4961.7
PubmedNMI—14.031.1435.2
PubmedARI—6.0—40.1
PubmedACC—53.060.8974.0
PubmedF1—17.060.1772.1
DBLPNMI63.261.056.2765.8
DBLPARI63.253.0—61.8
DBLPACC79.777.064.4978.6
DBLPF177.462.065.8278.5
20NewsNMI40.927.0—43.5
20NewsARI46.920.0—50.3
20NewsACC71.1858.0—76.0
20NewsF165.828.0—73.0

W2: Difference between HypeBoy and the k-means variant

Thank you for raising this point. Although the two methods share similar pretext tasks, they differ in training and implementation, with large performance gains observed mainly on Cora and 20News.

HypeBoy uses masked feature reconstruction (MFR) as warm-up, followed by hyperedge filling (HF); our variant jointly optimizes both losses. Their encoders and node representations for HF also differ. We use HypeBoy's published settings, selected by validation accuracy after node-classification fine-tuning (Appendix E.4), which may not suit clustering. Our variant also uses smaller learning rates and MFR/HF weights of 60/120. These differences may explain the performance gap.

The table below compares mean NMI and ACC (%). Large gaps occur only on Cora and 20News; differences on the other datasets are smaller, and HypeBoy performs better on Mushroom.

DatasetHypeBoy NMIk-means NMIHypeBoy ACCk-means ACC
Cora42.456.657.473.4
Citeseer40.444.766.868.1
Pubmed31.534.268.771.9
20News37.644.559.374.3
Mushroom44.742.586.585.8
Aminer42.143.433.136.3

Q1: Default parameter selection and learning rate

We selected a suitable default configuration through preliminary experiments on Cora, a small dataset, and transferred it to the other datasets. The default learning rate is 5×10−45\times10^{-4}. On Citeseer and DBLP, we observed unstable training at this rate and therefore reduced it to 10−410^{-4} to improve optimization stability, while retaining the other main parameter defaults. Thus, the dataset-specific learning rates reflect a stability adjustment to the shared configuration.

Q2: Per-epoch computational complexity

Let n,mn,m be the numbers of vertices and original hyperedges, I0I_0 the number of original incidences, and d,dzd,d_z the input and hidden dimensions. Topology enhancement retains at most rr vertices per induced hyperedge, giving at most I0+mrI_0+mr enhanced incidences. We assume fixed network/tree depths and internal cluster counts, and one full-batch update per epoch.

For relaxed HSE, the main overhead is sorting assignment values within each hyperedge to evaluate the Lovasz extension. Sorting only nonzero incidences costs O(∑e∣e∣log⁡∣e∣)O(\sum_e |e|\log|e|) over the enhanced hyperedges. Including assignment composition and the remaining evaluations, the HSE branch costs O(n+(I0+mr)log⁡n)O(n+(I_0+mr)\log n).

With exhaustive hyperedge filling, HSENet's per-epoch time complexity is

O(nddz+(n+m)dz2+I0d+n(I0+m)dz+(I0+mr)log⁡n).O\left(ndd_z+(n+m)d_z^2+I_0d+n(I_0+m)d_z+(I_0+mr)\log n\right).

For fixed feature dimensions, this simplifies to O(n(I0+m)+(I0+mr)log⁡n)O(n(I_0+m)+(I_0+mr)\log n). We use sampling on datasets with large hyperedge lengths and node numbers to reduce the filling cost from O(nI0dz)O(nI_0d_z) to O(qsdz)O(qsd_z), where qq is the number of queries and ss is the number of candidates per query. With fixed query budgets per hyperedge and candidate budgets, the overall bound becomes O(nm+(I0+mr)log⁡n)O(nm+(I_0+mr)\log n). Thus, the time complexity is dominated by topology enhancement, O(nm)O(nm), which is usually smaller than or comparable to O(n2)O(n^2). Details are in Appendix C.4.

Thank you for addressing my concerns. I will retain my rating for this submission.

Thank you for your thoughtful evaluation and continued support. We appreciate your time and consideration.

Reviewer_sFc1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HSENet for attributed hypergraph clustering. It formulates structural entropy for hypergraphs with edge-dependent vertex weights and uses a Lovász extension to support gradient-based soft partition learning. A shared hypergraph encoder learns from masked feature reconstruction and hyperedge filling. Clean representations induce additional weighted hyperedges, and an assignment network minimizes structural entropy on the enhanced hypergraph together with an orthogonality regularizer. Experiments on eight datasets compare against ten baselines, with the best mean results on all four metrics for five datasets.

Critical strengths
  1. The formulation connects hyperedge splitting costs to hierarchical assignment learning. The Lovász extension agrees with discrete cuts on binary assignments and supports almost-everywhere differentiable optimization. This provides a concrete way to incorporate high-order partition costs into representation learning.

  2. The architecture couples generative representation learning, topology enhancement, and direct cluster assignment in one objective. The encoder receives feedback from partition quality while retaining attribute and hyperedge-membership information. This addresses the separation between representation learning and clustering.

  3. The full system achieves substantial gains on several datasets. For example, ACC is 74.0 versus 69.4 for the best baseline on Pubmed, and 76.0 versus 71.2 on 20News. Five-run variability, component ablations, efficiency measurements, and attribute-masking experiments provide complementary evidence for the combined model.

Critical weaknesses
  1. The benefit of the hypergraph-specific objective remains unclear. Although Section 2 discusses LSEnet and DeSE, Table 1 includes no structural-entropy baseline on a documented graph expansion. A controlled graph-SE replacement using comparable representations and assignment capacity would help establish what HSE adds beyond existing structural-entropy clustering.

  2. The topology update can reinforce the errors already present in the embeddings. Each original hyperedge is reduced to one mean prototype, and new members are selected by cosine similarity (Section 5.2). For a hyperedge spanning several groups, this prototype can blur their differences; the resulting affinities then become evidence for the partition itself. The construction does not check whether the selected members form a coherent group. Appendix B.4 assumes a topology-quality condition that this selection rule does not enforce.

  3. The missing-attribute experiment only compares HSENet with its own ablations (Appendix E.4). It therefore does not show whether the advantage over competing methods in Table 1 survives incomplete inputs. On Mushroom, NMI falls from 0.4575 to 0.2051 at nominal 60% masking (Table 6). Without CAHC or HypeBoy under the same masks, the relative robustness of HSENet remains unclear.

Questions for authors
  1. Can you compare HSE with graph structural entropy on a stated expansion of the same hypergraph, keeping the representations and assignment capacity comparable? This would help establish the contribution of the high-order objective.

  2. How does topology enhancement handle an original hyperedge that mixes several groups? What allows an incorrect prototype or selected membership to be corrected, rather than reinforced through the HSE loss? Please relate the actual update rule to the topology-quality condition in Appendix B.4.

  3. Can you repeat the attribute-missing evaluation with CAHC and HypeBoy using the same fixed masks? Does HSENet retain its lead as masking increases, and how do the baselines compare on Mushroom, where its NMI drops by 55.2%?

5 comments
AuthorsOpenReview ↗

Response to W1/Q1 & W3/Q3

W1/Q1: Contribution of the hypergraph-specific objective

HSE improves average performance over a controlled Graph SE replacement, while HSENet outperforms the evaluated standalone graph-SE methods on all eight datasets.

We added two comparisons. Graph SE replacement changes only the structural objective within HSENet, retaining the encoder, assignment network, dimensions, hierarchy depth (K=2K=2), topology enhancement, auxiliary losses, and training settings. It evaluates the same soft hierarchy on a weighted clique expansion of the enhanced EDVW hypergraph:

A_{ii}=0,\qquad \Gamma_j=\sum_iH_{ij}.$$ The replacement evaluates the same soft hierarchy and uses the expanded graph only for loss evaluation. Thus, representation-learning and assignment capacities remain unchanged. **Standalone baselines** including SCISE, DeSE, LSEnet, and ASIL use binary clique expansions of the original hypergraphs, connecting vertices that share a hyperedge. Results below are percentages (mean±std, 5 runs); **bold marks each row’s highest mean**. OOM denotes out of memory. | Dataset | Metric | SCISE | DeSE | LSEnet | ASIL | Graph SE replacement | HSENet | |---|---|---|---|---|---|---|---| | Cora | NMI | 53.1±1.1 | 39.8±3.7 | 29.4±7.3 | 32.0±3.8 | 55.0±1.9 | **57.5±0.8** | | Cora | ACC | 71.7±0.8 | 58.9±6.8 | 41.0±11.2 | 45.6±2.6 | 70.5±4.8 | **74.7±1.6** | | Citeseer | NMI | 39.9±0.6 | 22.1±2.3 | 24.4±5.5 | 22.4±1.9 | 43.7±1.4 | **45.6±0.6** | | Citeseer | ACC | 64.1±0.9 | 47.7±4.0 | 46.7±7.8 | 45.9±3.5 | 69.5±2.0 | **72.1±0.4** | | Pubmed | NMI | 33.6±0.5 | 17.3±8.3 | 25.1±7.5 | 24.4±2.5 | 34.6±1.7 | **35.2±2.0** | | Pubmed | ACC | 72.8±0.4 | 57.9±8.9 | 63.7±8.6 | 55.6±0.6 | 72.5±3.2 | **74.0±2.0** | | DBLP | NMI | 59.4±2.9 | OOM | OOM | OOM | 65.7±2.2 | **65.8±2.6** | | DBLP | ACC | 72.3±6.1 | OOM | OOM | OOM | **79.0±5.3** | 78.6±5.5 | | 20News | NMI | 29.7±1.5 | OOM | OOM | OOM | **47.2±1.5** | 43.5±1.2 | | 20News | ACC | 61.6±1.2 | OOM | OOM | OOM | **77.9±2.0** | 76.0±1.0 | | Mushroom | NMI | 21.4±0.4 | OOM | OOM | OOM | 36.7±11.7 | **45.8±0.1** | | Mushroom | ACC | 75.9±0.2 | OOM | OOM | OOM | 80.6±10.0 | **86.8±0.0** | | AMiner | NMI | 42.9±0.3 | OOM | OOM | 45.9±4.9 | 47.2±1.3 | **49.9±0.9** | | AMiner | ACC | 35.7±0.6 | OOM | OOM | 41.1±3.7 | 41.5±1.7 | **43.1±0.6** | | Zoo | NMI | 72.8±1.9 | 0.0±0.0 | 14.1±0.6 | 10.9±0.7 | 87.3±2.5 | **90.5±1.9** | | Zoo | ACC | 73.3±3.3 | 40.6±0.0 | 45.0±0.8 | 31.5±0.5 | 86.1±2.9 | **93.1±0.6** | HSE improves average NMI/ACC over the controlled replacement by **2.04/2.59 percentage points**, winning on 7/6 datasets. This supports its contribution under comparable representation-learning and assignment capacities, although Graph SE leads on 20News and in DBLP ACC. HSENet also exceeds all available standalone baseline results in both metrics. Expansion definitions and detailed results have been added in Section 6.3 and Appendix E.1&2. ## W3/Q3: Robustness relative to CAHC and HypeBoy **HSENet retains an overall advantage on all 6 datasets other than Mushroom, achieving the highest ACC on each even at 60% attribute masking.** We evaluated CAHC and HypeBoy on the same seven datasets at $r\in\{0,0.1,\ldots,0.6\}$. For each ratio and seed, all methods share the same attribute masks, fixed throughout training. The table reports results at nominal **60% masking**, reporting with percentages (mean±std, 5 runs). | Dataset | Metric | HSENet | CAHC | HypeBoy | |---|---|---|---|---| | Cora | NMI | **50.75±1.60** | 49.07±1.75 | 37.96±1.73 | | Cora | ACC | **69.05±3.11** | 67.70±3.40 | 50.49±3.67 | | Citeseer | NMI | **40.07±1.07** | 37.30±0.84 | 35.38±1.42 | | Citeseer | ACC | **67.54±0.87** | 60.55±2.55 | 61.34±2.05 | | Pubmed | NMI | **35.11±1.06** | 32.14±1.04 | 30.42±1.25 | | Pubmed | ACC | **71.93±1.35** | 70.76±0.65 | 67.37±2.29 | | 20News | NMI | 36.13±2.02 | 26.79±4.11 | **37.12±0.57** | | 20News | ACC | **67.24±1.79** | 56.95±5.97 | 60.00±0.84 | | Mushroom | NMI | 20.51±6.78 | 37.19±3.53 | **41.48±1.30** | | Mushroom | ACC | 69.26±3.50 | 84.02±1.34 | **85.56±0.42** | | AMiner | NMI | **40.93±1.24** | 27.52±5.88 | 35.52±0.26 | | AMiner | ACC | **41.02±0.28** | 31.57±4.52 | 31.38±0.41 | | Zoo | NMI | **85.95±2.56** | 61.08±2.80 | 50.56±5.13 | | Zoo | ACC | **85.35±1.15** | 61.39±4.15 | 49.31±3.88 | HSENet leads in **NMI on five datasets and ACC on six** at 60% masking. The full curves (see revised manuscript) further show leading or competitive performance as missingness increases, with particularly stable performance on Pubmed and consistently strong results on Zoo. On 20News, its ACC exceeds HypeBoy by **7.24 points**, despite slightly lower NMI. Mushroom shows weaker missing-attribute robustness: HSENet’s NMI drops by **55.2%**, and both baselines outperform it under heavy masking. The shared-mask protocol, detailed results, and full curves have been added in Appendix E.6.
AuthorsOpenReview ↗

Response to W2/Q2

W2/Q2: Topology Enhancement for Mixed Hyperedges

Thank you for raising this question. We conduct 3 experiments to answer 3 sub-questions and show how topology enhancement refines mixed hyperedges.

1. How does topology enhancement handle an original hyperedge that mixes several groups?

Each source prototype selects up to rr nodes globally using positive cosine similarity, forming a potentially more coherent neighborhood. We ask what happens when an original hyperedge mixes several groups, and find Topology enhancement improves the average size-adjusted consistency of mixed-source hyperedges on all 8 datasets.

We measure purity as the fraction of nodes in an edge belonging to its majority class. Since purity depends on edge size, we also report excess purity XX: observed purity minus the expected purity of a uniformly random node set of the same size. Thus, ΔX=Xgen−Xsrc\Delta X=X_{\mathrm{gen}}-X_{\mathrm{src}} measures size-adjusted improvement, while the size effect accounts for the remaining raw purity change (ΔP=ΔX+size effect\Delta P=\Delta X+\text{size effect}).

The following results concern mixed-class source hyperedges

DatasetSource purityGenerated puritySize-adjusted gain ΔX\Delta XSize effect
Zoo.538.791.228±.009+.025
Cora.599.740.250±.002−.108
Citeseer.587.667.152±.002−.072
Pubmed.651.828.138±.001+.039
Mushroom.759.984.154±.008+.072
20News.723.913.055±.006+.134
Aminer.556.533.127±.001−.150
DBLP.654.779.193±.000−.069

Raw purity gain (ΔP\Delta P: Generated - Source) is positive on 7 datasets; size-adjusted gains (ΔX\Delta X) are positive on all 8 (.055–.250), also in training-mode samples. Cora improves from .599 to .740 despite an adverse size effect. These results verify that topology enhancement improves mixed hyperedges.

2. What allows an incorrect prototype or selected membership to be corrected, rather than reinforced through the HSE loss?

Each forward pass recomputes prototypes from original members' embeddings and reselects memberships. Encoder updates from clustering and generative supervision allow early selections to change. We ask whether incorrect prototypes or memberships can change rather than be reinforced, and find Initially low-quality edges improve on average on all 8 datasets.

Immediately before HSE optimization, we identify nonempty generated edges whose purity does not exceed the size-matched random expectation, then track the same edges until training ends. “Improved” and “deteriorated” refer to increases and decreases in purity.

DatasetSelected edgesImprovedDeterioratedMean purity gainExcess purity XX: start → end
DBLP25.896.7%0%+.300±.021−.118 → +.182
Aminer22.689.9%0%+.129±.015−.081 → +.048
Pubmed64.681.4%0%+.345±.021−.316 → +.029
Zoo14.068.5%4.1%+.234±.025−.041 → +.193
Citeseer98.440.5%0.2%+.096±.005−.046 → +.050
Cora18.829.8%0%+.056±.008−.106 → −.049
Mushroom1.8100%0%+.280±.045−.084 → +.196
20News1.462.5%0%+.196±.213−.045 → +.151

Improvements are seen on most hyperedges, while deterioration is barely seen. Mean purity rises by .056–.345 across all 8 datasets. These results verify that low-quality hyperedges improve during training.

3. How does the actual update relate to Appendix B.4's condition?

On 7 of 8 datasets, generated and fused topologies reduce conductance.

For a fixed cut SS, let ϕq=gq/dq\phi_q=g_q/d_q, where gqg_q is cut cost and dqd_q the smaller-side volume (oo: original; rr: generated). For positive denominators, B.4 gives

ϕβ(S)≤doϕo+βdrϕrdo+βdr.\phi_\beta(S)\le\frac{d_o\phi_o+\beta d_r\phi_r}{d_o+\beta d_r}.

Thus ϕr≤ϕo\phi_r\le\phi_o ensures non-worsening fusion. We test fixed true-class cuts independently of predictions, using actual incidence weights and β=3\beta=3.

Table 3. Conductance under linear cuts. k/Ck/C reports the mean number of class cuts with ϕr<ϕo\phi_r<\phi_o across seeds (kk), out of CC evaluated cuts.

Dataset# ϕr<ϕo\phi_r<\phi_o: mean k/Ck/CMedian ϕr−ϕo\phi_r-\phi_oMedian ϕβ−ϕo\phi_\beta-\phi_o
Zoo5/7−.531−.263
Cora7/7−.042−.032
Citeseer3.6/6−.003−.002
Pubmed3/3−.048−.025
Mushroom1/1−.116−.010
20News4/4−.494−.008
Aminer0/12+.161+.126
DBLP4.6/6−.004−.003

On 7 datasets, more than half of the class cuts show lower conductance for both the generated and fused topologies, averaged across seeds. Median conductance changes are also negative for both topologies on these 7 datasets. Thus, the learned updates satisfy B.4's quality condition on most evaluated cuts in 7 datasets, and fusion improves those cuts as predicted.

AuthorsOpenReview ↗

Response to W1&Q1: Detailed analysis of contribution of the hypergraph-specific objective

W1/Q1: Contribution of the hypergraph-specific objective

Thank you for this helpful suggestion. We expand on our earlier brief response below, detailing the experiments in Section 6.3 and Appendices E.1–E.2 of the revised PDF previously referenced. We conduct two comparisons to evaluate existing graph-SE methods and isolate HSE's contribution under matched model capacity.

Standalone graph-SE baselines. We compare HSENet with SCISE, DeSE, LSEnet, and ASIL on the same eight datasets. These methods use binary clique expansions of the original hypergraphs: distinct vertices are connected if they share a hyperedge. HSENet operates directly on hypergraphs. This comparison evaluates the complete clustering methods.

Controlled objective replacement. We replace only the HSE loss within HSENet with graph structural entropy (G-SE), retaining the hypergraph encoder, assignment network, representation dimensions, hierarchy depth (K=2K=2), topology-enhancement procedure, auxiliary losses, and training settings. At each training step, G-SE uses a weighted clique expansion of the enhanced EDVW hypergraph: Aiℓ=∑jω(ej)ΓjHijHℓj(i≠ℓ),Aii=0,Γj=∑iHij,A_{i\ell}=\sum_j\frac{\omega(e_j)}{\Gamma_j}H_{ij}H_{\ell j}\quad(i\ne\ell),\qquad A_{ii}=0,\quad\Gamma_j=\sum_iH_{ij}, where HH is the enhanced incidence matrix and ω(ej)\omega(e_j) is the hyperedge weight. Graph degrees and volumes are recomputed after removing self-loops. The graph is used only to evaluate the graph-SE loss on the existing soft hierarchy, including vertex leaves. Thus, only the structural objective changes, with representation-learning and assignment capacities preserved.

Results are percentages (mean±std, 5 runs). G-SE is the controlled replacement; HSE is full HSENet. F1 denotes Macro-F1; OOM means out of memory. Bold marks the highest mean in each row.

DatasetMetricSCISEDeSELSEnetASILG-SEHSE
CoraNMI53.1±1.139.8±3.729.4±7.332.0±3.855.0±1.957.5±0.8
ARI48.9±1.134.4±5.021.7±9.424.4±3.349.8±3.552.9±0.9
ACC71.7±0.858.9±6.841.0±11.245.6±2.670.5±4.874.7±1.6
F170.1±0.946.6±10.821.3±13.627.3±3.366.9±5.169.5±4.0
CiteseerNMI39.9±0.622.1±2.324.4±5.522.4±1.943.7±1.445.6±0.6
ARI39.2±0.718.2±2.521.8±6.420.9±1.844.9±2.447.9±0.7
ACC64.1±0.947.7±4.046.7±7.845.9±3.569.5±2.072.1±0.4
F159.6±1.040.6±3.834.2±11.131.1±5.159.5±1.661.7±0.4
PubmedNMI33.6±0.517.3±8.325.1±7.524.4±2.534.6±1.735.2±2.0
ARI38.4±1.015.8±7.925.5±9.822.9±3.638.0±4.040.1±3.0
ACC72.8±0.457.9±8.963.7±8.655.6±0.672.5±3.274.0±2.0
F170.9±0.341.8±8.551.6±11.145.4±1.668.1±8.472.1±1.8
DBLPNMI59.4±2.9OOMOOMOOM65.7±2.265.8±2.6
ARI56.3±4.5OOMOOMOOM63.3±4.261.8±5.1
ACC72.3±6.1OOMOOMOOM79.0±5.378.6±5.5
F169.9±6.5OOMOOMOOM76.9±7.178.5±4.9
20NewsNMI29.7±1.5OOMOOMOOM47.2±1.543.5±1.2
ARI30.8±2.0OOMOOMOOM53.7±2.050.3±1.2
ACC61.6±1.2OOMOOMOOM77.9±2.076.0±1.0
F153.9±1.6OOMOOMOOM75.2±2.573.0±1.7
MushroomNMI21.4±0.4OOMOOMOOM36.7±11.745.8±0.1
ARI26.8±0.4OOMOOMOOM41.4±18.754.2±0.1
ACC75.9±0.2OOMOOMOOM80.6±10.086.8±0.0
F175.8±0.2OOMOOMOOM79.4±12.186.7±0.0
AMinerNMI42.9±0.3OOMOOM45.9±4.947.2±1.349.9±0.9
ARI28.1±0.7OOMOOM31.2±4.633.4±2.135.4±0.5
ACC35.7±0.6OOMOOM41.1±3.741.5±1.743.1±0.6
F127.3±0.5OOMOOM17.7±2.724.0±1.322.1±1.4
ZooNMI72.8±1.90.0±0.014.1±0.610.9±0.787.3±2.590.5±1.9
ARI64.1±4.10.0±0.04.1±0.32.3±0.585.4±4.591.8±3.2
ACC73.3±3.340.6±0.045.0±0.831.5±0.586.1±2.993.1±0.6
F163.0±2.38.3±0.022.4±2.418.9±0.266.5±6.978.7±1.8

Comparison with standalone methods. HSENet outperforms every standalone graph-SE baseline with available results in NMI, ARI, and ACC on all eight datasets, and achieves the highest Macro-F1 on six. Its NMI gains over the strongest available standalone baseline reach 24.4 points on Mushroom and 13.8 points on 20News, demonstrating its advantage over these graph-based clustering pipelines.

Contribution of HSE under matched capacity. HSE shows a clear overall advantage over G-SE across the eight datasets, improving average NMI/ARI/ACC/Macro-F1 by 2.05/3.06/2.60/3.23 percentage points, respectively. It achieves higher means in 25 of 32 dataset–metric pairs, with gains across all four metrics on five datasets. These improvements arise with the same model architecture and assignment capacity, supporting HSE as an important contributor to HSENet's performance. Directly modeling hyperedge splitting costs, which clique expansion does not generally preserve, may account for this advantage.

The expansion and objective definitions are in Appendix D.4, with complete results in Appendices E.1–E.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.