acceptodds
Under review as a conference paper at ICLR 2027

AD-DHGNN: Adaptive Dual-Controlled Dynamic Hypergraph Neural Networks

11 views

Abstract

Dynamic hypergraph neural networks adapt high-order relations to evolving node representations, yet most existing constructions rely on fixed hyperedge counts or hyperedge cardinality constraints. Such fixed structural choices can produce redundant, oversized, or fragmented hypergraphs when the required relation structure varies across layers and datasets. We propose the Adaptive Dual-Controlled Dynamic Hypergraph Neural Network AD-DHGNN, which treats global hyperedge count and local hyperedge cardinality as two complementary structural variables. AD-DHGNN first generates posterior-weighted candidate hyperedges from anchor-based node similarities, avoiding dense node-pair affinity estimation. It then applies a cardinality controller that assigns data-dependent capacity budgets to individual candidates, followed by a count controller that selects a compact and representative subset of hyperedges. The resulting sparse hypergraph is used for gated dynamic message passing. Experiments on seven node-classification benchmarks show that AD-DHGNN improves predictive performance over the evaluated baselines while constructing compact and low-redundancy dynamic hypergraphs. These results demonstrate that explicitly determining hyperedge count and hyperedge cardinality is a central principle for scalable dynamic hypergraph reconstruction.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic hypergraph neural networks, Adaptive hyperedge cardinality control, Adaptive hyperedge count control, Higher-order representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_rALFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes AD-DHGNN, a dynamic hypergraph neural network that adaptively determines both the number of hyperedges and the number of nodes within each hyperedge during encoding.

It combines a PageRank-based graph encoder with a Gaussian mixture–based candidate generation pipeline and adaptive structural controllers.

Experiments compare the proposed method against several hypergraph neural network baselines on seven node-classification benchmarks.

Critical strengths
  • S1. The key challenge, selecting informative hyperedges, is an important and practical research question for the use cases of hypergraph neural networks.

  • S2. Adaptive hyperedge selection specialized to each dataset is reasonable and intuitive.

Critical weaknesses
  • W1. Methodological novelty. The proposed framework combines established techniques, including APPNP, Gaussian mixture modeling, and spectral knee detection. Although their integration for joint hyperedge count and cardinality control is potentially useful, the methodological advance beyond this combination remains unclear. The authors should clarify which design elements introduce new insights specific to hypergraph construction and how they differ from existing adaptive structure-learning approaches.

  • W2. Baseline coverage. Although the experiments include recent hypergraph methods, several relevant comparisons are missing. For example, FDHNN (2026) is discussed in the related work but is not evaluated. Moreover, because the evaluated datasets are originally graphs and the proposed method uses an APPNP backbone, comparisons with competitive graph neural networks would help establish the benefits of hypergraph reconstruction. The authors should expand the baseline coverage or explain the exclusions.

  • W3. Relationship to homophily. The method constructs hyperedges from representation similarity, making its relationship to homophily particularly relevant. However, the paper does not systematically examine whether the reconstructed hyperedges exhibit greater label homophily or how their homophily relates to classification performance. Evaluating datasets with varying levels of homophily would also help clarify whether the method remains effective when similar or connected nodes have different labels.

Questions for authors

See Weakness

5 comments

Response to W1. Thank you for the reviewer’s suggestion. Existing dynamic hypergraph methods update hyperedge based on node representations, and some also adjust the number of hyperedges. Our design further addresses two decisions in hypergraph construction.

First, different hyperedges may require different cardinality. A dense local region may support a larger hyperedge cardinality, whereas a small local group should not be forced to expand. We therefore use the size of each candidate’s local region to set its cardinality upper bound, then select members according to the strength of their association with that candidate. The local region determines how many nodes the candidate can contain; it does not directly determine which nodes it must contain.

Second, the model must determine both how many hyperedges to retain and which ones to keep. Similarity among candidates helps identify repeated structure, but using similarity alone may remove candidates that cover different nodes or regions. We therefore combine structural redundancy with complementary coverage to guide both the retained count and candidate selection.

Together, these two designs limit the overexpansion of individual hyperedges and reduce duplication among candidates. APPNP provides node representations informed by the original graph, the GMM estimates candidate membership, and spectral knee detection informs the number of hyperedges to retain.

The proposed method focuses on assigning candidate-specific size limits, rather than using k-means or k-nearest neighbours, and selecting the final hyperedges based on both structural redundancy and complementary coverage. As shown in Table 1, AD-DHGNN produces a more balanced hyperedge count and cardinality range than the compared methods, with substantially lower structural redundancy.

Table 1. Structural properties of the constructed hypergraphs.

DatasetMethodHyperedge countCardinality rangeJaccard redundancySpectral redundancy
CoraDHGNN8,124[1, 98]0.72370.8970
CoraDHKH8,124[2, 98]0.48440.8663
CoraTDHNN17[163, 2708]0.72410.7689
CoraDHHNN13[1, 2708]0.77000.8285
CoraAD-DHGNN22[15, 134]0.14580.0509
PubMedDHGNN59,151[1, 92]0.46680.8742
PubMedDHKH59,151[2, 92]0.32460.8278
PubMedTDHNN11[5,639, 19717]0.90910.8765
PubMedDHHNN12[197, 19717]0.83340.8492
PubMedAD-DHGNN132[2, 441]0.13260.0613
PhysicsDHGNN103,479[1, 108]0.48620.8678
PhysicsDHKH103,479[2, 108]0.37550.8270
PhysicsTDHNN50[1, 29515]0.75170.7347
PhysicsDHHNN13[26, 34493]0.77030.8300
PhysicsAD-DHGNN48[18, 4952]0.11610.0558

Response to W2. Thank you for the reviewer’s comment. We evaluated FDHNN on all seven real-world datasets. We also compared AD-DHGNN with validation-tuned GCN, GAT, and APPNP models using the same dataset splits. All experiments were conducted on an RTX A800 GPU with 80 GB of memory. Table 1 shows the baseline hyperparameter settings.

Table 1. Baseline hyperparameter settings.

MethodHyperparameter settings
GCNHidden dimension: {64, 128, 256}. Layers: {2, 3}. Dropout: {0, 0.3, 0.5, 0.7}. Learning rate: {0.001, 0.003, 0.01}. Weight decay: {0, 0.0005, 0.005}.
GATHidden dimension, dropout, learning rate, and weight decay use the GCN ranges. Attention heads: {2, 4, 8}. Attention dropout: {0, 0.3, 0.5, 0.7}.
APPNPHidden dimension, dropout, learning rate, and weight decay use the GCN ranges. Propagation steps K: {5, 10, 20}. Teleport probability α: {0.05, 0.1, 0.2}.
FDHNNHidden dimension: 128. Number of edges: {100, 150, 200}. Minimum number of edges: {10, 20, 30}. k: 10. Learning rate: 0.001. Weight decay: 0.

Table 2 reports the results. AD-DHGNN outperforms FDHNN on all five datasets where FDHNN completed its run. FDHNN controls hyperedge cardinality and count through preset hyperparameters, whereas AD-DHGNN adapts both using data-dependent structural information. This difference may partly explain AD-DHGNN’s stronger classification performance.

CiteSeer is the only dataset on which AD-DHGNN does not exceed GAT. CiteSeer has relatively sparse connections and high-dimensional features. These dataset characteristics may make node–anchor similarity estimates less stable and contribute to the smaller result on CiteSeer.

Table 2. Classification performance (%). The first six datasets use accuracy, and Yelp uses Micro-F1. Results are mean ± standard deviation.

MethodCoraCiteSeerPubMedPhysicsAmazon Computersogbn-arxivYelp
GCN81.56 ± 0.6171.86 ± 0.2877.94 ± 0.4696.70 ± 0.0769.22 ± 0.7865.69 ± 0.2222.09 ± 3.26
GAT81.46 ± 0.5071.50 ± 0.7677.48 ± 0.4796.45 ± 0.0571.85 ± 3.6365.36 ± 0.0720.37 ± 3.05
APPNP82.72 ± 0.4171.10 ± 0.5578.42 ± 0.2196.60 ± 0.0669.68 ± 0.5265.55 ± 0.2421.52 ± 1.88
FDHNN79.40 ± 0.5964.36 ± 1.3876.94 ± 1.0596.30 ± 0.7871.22 ± 0.86OOMOOM
TDHNN82.77 ± 0.3265.81 ± 1.2378.13 ± 0.3796.52 ± 0.1869.97 ± 0.9352.77 ± 0.8232.20 ± 1.15
AD-DHGNN83.57 ± 0.1371.22 ± 0.2480.14 ± 0.1797.26 ± 0.1180.72 ± 0.7170.35 ± 0.2350.83 ± 0.17

Response to W3. Thank you for raising this question. We measure hyperedge label homophily as the proportion of same-label node pairs among all node pairs within the evaluated hyperedges:

\frac{ \sum_{e\in\mathcal E} \sum_{i < j,\\,v_i,v_j\in e}\mathbf{1}[y_i=y_j] }{ \sum_{e\in\mathcal E}\binom{|e|}{2} }.$$ Here, $\mathcal E$ is the evaluated hyperedge set and $y_i$ is the label of node $v_i$. Labels are used only for post-inference analysis, not for hyperedge construction. Table 1 compares the label homophily and classification accuracy obtained by using candidate hyperedges directly with those obtained after adaptive dual-control reconstruction. The reconstructed hyperedges have higher label homophily, and the model achieves higher classification accuracy. **Table 1. Effect of adaptive dual-control reconstruction on hyperedge label homophily and classification accuracy on Cora.** | Hyperedge type | Label homophily (%) | ACC (%) | |---|---:|---:| | Before Reconstructed | 31.30 ± 2.11 | 81.67 ± 1.46 | | Reconstructed Hyperedges | **86.50 ± 0.96** | **83.57 ± 0.13** | To compare performance across different homophily levels, we use **original-graph edge homophily** as the proportion of edges connecting same-label nodes in the undirected input graph, excluding self-loops: $$h_{\mathrm{graph}} = \frac{ \sum_{\\{i,j\\}\in E,\\,i\ne j} \mathbf{1}[y_i=y_j] }{ \left|\\{\\{i,j\\}\in E \mid i\ne j\\}\right| }.$$ Here, $E$ is the set of undirected edges in the input graph, $y_i$ is the label of node $i$, $i \ne j$ excludes self-loops, and $\mathbf{1}[y_i=y_j]$ equals 1 if nodes $i$ and $j$ share the same label and 0 otherwise. Table 2 presents the corresponding results. The original-graph edge homophily of Roman-empire, ogbn-arxiv, and Physics is 4.69%, 65.42%, and 93.14%, respectively. AD-DHGNN exceeds GCN in test accuracy by 7.93, 4.66, and 0.56 percentage points on these datasets. The results show classification gains in the evaluated low, medium, and high-homophily settings. AD-DHGNN achieves gains at all three homophily levels, with the largest improvement over GCN on the low-homophily Roman-empire dataset. **Table 2. Classification accuracy on datasets with different levels of original-graph edge homophily.** | Dataset | Original-graph edge homophily (%) | GCN (%) | AD-DHGNN (%) | |---|---:|---:|---:| | Roman-empire | 4.69 (low) | 46.63 ± 0.17 | **54.56 ± 0.60** | | ogbn-arxiv | 65.42 (medium) | 65.69 ± 0.22 | **70.35 ± 0.23** | | Physics | 93.14 (high) | 96.70 ± 0.07 | **97.26 ± 0.11** |
Reviewer_VyebOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AD-DHGNN, a dynamic hypergraph network for node classification on graphs. The K highest-degree nodes serve as anchors. At each dynamic layer, a twocomponent Gaussian mixture is fitted to each anchor’s similarities with the other nodes, and the nodes with posterior above 1/2 (plus the anchor) form a candidate hyperedge. Two controllers then shape the hypergraph: a cardinality controller caps each candidate at the size of its anchor’s Voronoi region, and a count controller picks how many candidates to keep from the knees of a spectral curve and a greedy coverage curve. The kept hyperedges are used for gated message passing on top of an APPNP encoder. Results are reported on seven node classification benchmarks, with ablations of the two controllers on Cora.

Critical strengths
  1. The paper handles two different problems of learned hypergraphs with separate mechanisms: too many overlapping hyperedges, and single hyperedges that contain too many nodes. Eq. (6) bounds the size of each candidate, and Eqs. (7)–(12) decide how many candidates to keep. I think this split makes sense, because keeping fewer hyperedges does not by itself stop the remaining ones from growing very large.

  2. The Cora ablation (Table 3) supports the claim that the two controllers do different jobs. Count-only still has hyperedges with 1,595/1,412 nodes (layer 1/2) and Cardinality-only keeps all 64 candidates, while with both controllers the model keeps 22/23 hyperedges of at most 131/132 nodes. Accuracy is also highest with both (83.57% vs. 82.00% and 82.73%), although the gap to Cardinality-only is within its standard deviation.

Critical weaknesses
  1. The count rule can drop candidates that are almost disjoint from the ones it keeps. For two 100-node candidates sharing one node, Eq. (7) gives eigenvalues 1±ϵ, hence Mspectral = 1; the first greedy coverage gain is above 1/2, so Mcoverage = 1 too, and Eq. (12) keeps one hyperedge, leaving the other 99 nodes out. Had the candidates been disjoint, G would be the identity and both would be kept (Appendix A.7), so a single shared node flips the outcome. I don’t know how often this happens on real data (Q1), but the (1 − 1/e) guarantee in Sec. 3.3 assumes a fixed budget and does not address how it is chosen.

  2. There are no graph-only baselines in Table 2. APPNP-only is in Table 3 (Cora) and Table 7 (all datasets), but these numbers look under-tuned: 79.80% on Cora, 65.56% on CiteSeer, 34.59 ± 24.18% on Amazon Computers. For reference, the GCN and GAT papers report 81.5% and 83.0% on the public Cora split, and the OGB paper reports 71.74% for GCN on ogbn-arxiv, above the 70.35% of AD-DHGNN. So it is hard to tell how much the dynamic hypergraph adds over a well-tuned graph model.

  3. The ablation variants differ in the controllers and in how much structure they keep (about 6 large hyperedges for Count-only, 64 for Cardinality-only, 22/23 smaller ones for Dualcontrol), so Table 3 cannot separate adaptivity from keeping less structure. Appendix E.4 tests a uniform cap, but only on synthetic data, and there is no fixed-count baseline.

Questions for authors
  1. How often does the count rule drop candidates covering many nodes that no kept hyperedge covers? Could you report this (e.g. the number of such nodes) on the real datasets?

  2. How were the input hypergraphs for the static baselines built, and how were they tuned? How does AD-DHGNN compare with tuned GCN, GAT or APPNP on the same splits?

  3. At a matched number of hyperedges and memberships, do the adaptive controllers still beat fixed limits (fixed count, uniform size cap) on the real datasets?

5 comments

Response to Q1.

Thank you for the suggestion. We measured node coverage before and after hyperedge count control. A node is counted as losing coverage if it belongs to at least one candidate hyperedge but to no retained hyperedge. The coverage loss rate uses the number of distinct nodes covered by candidates before count control as its denominator.

Let CC and SS denote the sets of candidate and retained hyperedges, respectively. For any hyperedge set AA, let U(A)=⋃e∈AeU(A)=\bigcup_{e\in A}e denote the set of nodes it covers. The number of distinct nodes that lose coverage and the corresponding coverage loss rate are

\qquad \text{Coverage loss rate (\\%)}=100\\,\frac{N_{\mathrm{lost}}}{|U(C)|}.$$ As shown in Table 1, coverage loss occurs on all four evaluated datasets. The numbers of distinct nodes losing coverage are 241 (17.54%) on Cora, 1,125 (12.81%) on PubMed, 302 (1.61%) on Physics, and 3,690 (3.86%) on ogbn-arxiv. Thus, reducing the hyperedge count can leave some previously covered nodes outside the retained hyperedges, with the loss rate varying across datasets. **Table 1. Node coverage changes after hyperedge count control.** | Dataset | Candidate hyperedges | Deleted hyperedges | Retained hyperedges | Distinct nodes losing coverage (loss rate) | |---|---:|---:|---:|---:| | Cora | 53 | 36 | 17 | **241 (17.54%)** | | PubMed | 296 | 208 | 88 | **1,125 (12.81%)** | | Physics | 684 | 450 | 234 | **302 (1.61%)** | | ogbn-arxiv | 423 | 309 | 114 | **3,690 (3.86%)** |

Response to Q2. Thank you for the comment.

We constructed the input hypergraphs for the static hypergraph baselines following their publicly available implementations and selected construction parameters using validation results. For HGNN, we used KNN construction, searched k∈5,10,15,…,100k\in\\{5,10,15,\ldots,100\\}, and selected kk based on validation accuracy.

We also compared AD-DHGNN with GCN, GAT, and APPNP on the same dataset splits. Their hyperparameter search ranges are listed in Table 1, with configurations selected using validation performance. Table 2 shows that AD-DHGNN achieves a higher mean result than GCN on six of seven datasets, GAT on six of seven, and APPNP on all seven.

On CiteSeer, AD-DHGNN trails GCN by 0.64% and GAT by 0.28%. CiteSeer has relatively sparse connections and high-dimensional node features. These characteristics may make node–anchor similarity estimates less stable and limit the benefit of dynamic hyperedge reconstruction on this dataset.

Table 1. Hyperparameter search ranges for the graph baselines.

MethodHyperparameter settings
GCNHidden dimension ∈64,128,256\in\\{64,128,256\\}, layers ∈2,3\in\\{2,3\\}, dropout ∈0,0.3,0.5,0.7\in\\{0,0.3,0.5,0.7\\}, learning rate ∈0.001,0.003,0.01\in\\{0.001,0.003,0.01\\}, weight decay ∈0,0.0005,0.005\in\\{0,0.0005,0.005\\}
GATHidden dimension, dropout, learning rate, and weight decay use the GCN ranges. Attention heads ∈2,4,8\in\\{2,4,8\\}, attention dropout ∈0,0.3,0.5,0.7\in\\{0,0.3,0.5,0.7\\}
APPNPHidden dimension, dropout, learning rate, and weight decay use the GCN ranges. Propagation steps K∈5,10,20K\in\\{5,10,20\\}, teleport probability α∈0.05,0.1,0.2\alpha\in\\{0.05,0.1,0.2\\}

Table 2. Node-classification performance (%).

MethodCoraCiteSeerPubMedPhysicsAmazon Computersogbn-arxivYelp
GCN(Tuned)81.56 ± 0.6171.86 ± 0.2877.94 ± 0.4696.70 ± 0.0769.22 ± 0.7865.69 ± 0.2222.09 ± 3.26
GAT(Tuned)81.46 ± 0.5071.50 ± 0.7677.48 ± 0.4796.45 ± 0.0571.85 ± 3.6365.36 ± 0.0720.37 ± 3.05
APPNP(Tuned)82.72 ± 0.4171.10 ± 0.5578.42 ± 0.2196.60 ± 0.0669.68 ± 0.5265.55 ± 0.2421.52 ± 1.88
AD-DHGNN83.57 ± 0.1371.22 ± 0.2480.14 ± 0.1797.26 ± 0.1180.72 ± 0.7170.35 ± 0.2350.83 ± 0.17

Response to Q3. Thank you for the question. We evaluated fixed hyperedge counts MM and fixed cardinality caps LL on Cora, PubMed, and Physics. Tables 1 and 2 report the complete sweeps. We then combined the best-performing MM and LL from the two separate sweeps for each dataset. Table 3 shows that AD-DHGNN outperforms these fixed combinations on all three datasets by 1.87, 1.77, and 0.207 percentage points, respectively. AD-DHGNN also has lower standard deviations on Cora and PubMed, while the fixed combination has a lower standard deviation on Physics. Some individual fixed-count or uniform-cap settings come close to AD-DHGNN, but combining the two independently selected limits yields lower accuracy on all three datasets. The fixed approach requires tuning MM and LL for each dataset, while AD-DHGNN determines both during reconstruction. This result supports coordinating hyperedge count and cardinality during reconstruction.

Table 1. Test accuracy (%) with a fixed number of hyperedges per layer. Values are mean ± standard deviation. “—” indicates a setting that was not evaluated.

Fixed MM per layerCoraPubMedPhysics
1083.50 ± 1.80——
1582.97 ± 1.66——
2081.87 ± 0.91——
3082.40 ± 1.35——
4082.31 ± 1.1478.63 ± 0.85—
60—78.87 ± 1.00—
80—79.07 ± 1.39—
100—78.10 ± 0.4097.078 ± 0.13
150—78.47 ± 0.9597.124 ± 0.27
200——97.124 ± 0.11
250——97.147 ± 0.32
300——97.101 ± 0.36

Table 2. Test accuracy (%) with a uniform hyperedge cardinality cap.

Fixed LLCoraPubMedPhysics
380.90 ± 0.4078.70 ± 0.7297.113 ± 0.21
581.57 ± 1.5278.80 ± 0.7297.113 ± 0.11
1081.90 ± 1.3778.80 ± 0.7897.217 ± 0.37
2082.33 ± 0.9578.50 ± 0.3097.205 ± 0.42
3082.83 ± 1.7078.33 ± 1.0797.09 ± 0.31
5082.23 ± 0.5177.87 ± 0.9197.043 ± 0.25
7581.13 ± 0.2578.20 ± 0.8597.09 ± 0.44
10082.70 ± 0.8278.27 ± 0.5197.008 ± 0.20
15081.23 ± 0.8578.10 ± 1.0196.997 ± 0.23
20081.63 ± 1.1278.43 ± 0.6197.02 ± 0.57

Table 3. Test accuracy (%) with the selected fixed limits and adaptive dual control.

DatasetFixed MMFixed LLFixed combinationAD-DHGNN
Cora103081.70 ± 1.3283.57 ± 0.13
PubMed801078.37 ± 0.2380.14 ± 0.17
Physics2501097.055 ± 0.04297.262 ± 0.11

Comments on the paper

6 comments

Response to Q1.

Thank you for the suggestion. We evaluated single- and two-Gaussian fits to node–anchor similarity distributions at the second dynamic layer on all seven real datasets. For nn evaluated similarities s1,…,sns_{1}, \ldots, s_{n}, let F^\widehat{F} denote their empirical cumulative distribution function (CDF), and let FmF_{m} and fmf_{m} denote the fitted CDF and probability density of model mm, respectively. We compute

KSm=sup⁡t∣F^(t)−Fm(t)∣∈[0,1].\mathrm{KS}_{m} = \sup_{t} \lvert \widehat{F}(t) - F_{m}(t) \rvert \in [0,1].

NLLm=−1n∑i=1nlog⁡fm(si)∈R.\mathrm{NLL}_{m} = -\frac{1}{n}\sum_{i=1}^{n}\log f_{m}(s_{i}) \in \mathbb{R}.

Lower values indicate better fit. The KS statistic lies between 0 and 1. NLL can take positive or negative real values because a continuous probability density can exceed 1. In particular, NLL is negative when the average log density is positive.

As shown in Table 1, at the second dynamic layer, the two-Gaussian model achieves a lower NLL on all seven datasets and a lower KS statistic on six. Yelp is the only KS exception, with values of 0.4723 for the single-Gaussian fit and 0.4727 for the two-Gaussian fit. Yelp is the largest and only multi-label dataset in this comparison. Its node–anchor similarity distributions may therefore be more complex, limiting the improvement in the maximum CDF discrepancy even though the two-Gaussian model reduces NLL.

The results show that the two-Gaussian model fits the observed similarity distributions more closely than the single-Gaussian model on most datasets, with lower KS statistics on six datasets. On Yelp, however, the KS statistics of the two models are similar and remain high, indicating a limitation of the two-Gaussian approximation on this dataset. Thus, a two-Gaussian model cannot adequately describe every similarity distribution, and we will explore alternative distributions in future work.

Table 1. Goodness-of-fit diagnostics for node–anchor similarity distributions at the second dynamic layer.

DatasetKS: 1 Gaussian ↓KS: 2 Gaussians ↓NLL: 1 Gaussian ↓NLL: 2 Gaussians ↓
Cora0.08160.04340.11300.0302
CiteSeer0.08100.04600.28600.1976
PubMed0.13960.06180.41560.1488
Physics0.26010.11290.4693−0.6356
Amazon Computers0.10690.04780.23210.0519
ogbn-arxiv0.12440.04010.19580.0303
Yelp0.47230.4727−3.6000−3.6489

Response to Q2

Thank you for the question. We conducted a sensitivity analysis on Cora, varying one parameter at a time and keeping the others at their default values. Test accuracy is reported as mean ± standard deviation. As shown in Tables 1–4, mean accuracy is relatively insensitive to the tested changes in EM initialization, iteration count, and posterior threshold. EM initialization has a more pronounced effect on variability across runs. Among the tested settings, raising the variance floor to 10−310^{-3} has the largest effect on both mean accuracy and variability across runs. The higher floor constrains the fitted Gaussian variances to different degrees across runs, causing nodes with posterior probabilities near the threshold to enter candidate hyperedges in some runs and be excluded in others, which may increase variability in classification accuracy.

Table 1. EM mean initialization.

Initial component meansTest accuracy (%)
μ±0.1σ\mu \pm 0.1\sigma83.03 ± 0.90
μ±0.5σ\mu \pm 0.5\sigma (default)83.57 ± 0.13
μ±1.0σ\mu \pm 1.0\sigma83.27 ± 0.76

Table 2. Number of EM iterations.

EM iterationsTest accuracy (%)
483.23 ± 0.42
8 (default)83.57 ± 0.13
1683.67 ± 0.15

Table 3. Variance floor.

Variance floorTest accuracy (%)
10−510^{-5}83.50 ± 0.17
10−410^{-4} (default)83.57 ± 0.13
10−310^{-3}82.50 ± 1.06

Table 4. Posterior membership threshold.

Posterior thresholdTest accuracy (%)
0.383.17 ± 0.45
0.5 (default)83.57 ± 0.13
0.783.60 ± 0.20

Response to Q3. Thank you for the suggestion. We compared five anchor selection strategies on Cora, PubMed, and Physics: random selection, k-center, k-means, jointly learned selection, and selection by highest node degree (AD-DHGNN). In the jointly learned variant, we train a node scoring module together with the node classification model and use the learned scores to select anchors from the existing graph nodes.

As shown in Table 1, AD-DHGNN’s anchor selection strategy achieves the highest mean accuracy on all three datasets, exceeding the strongest alternative by 0.40, 2.00, and 0.50 percentage points on Cora, PubMed, and Physics, respectively. The higher mean accuracy support AD-DHGNN’s anchor selection strategy as a reliable for candidate hyperedge generation on the evaluated datasets.

Table 1. Node classification accuracy (%) under different anchor selection strategies.

Anchor selectionCoraPubMedPhysics
Random80.27 ± 0.3177.82 ± 0.3296.76 ± 0.31
K-center82.62 ± 0.4678.12 ± 0.5596.64 ± 0.29
K-means83.17 ± 0.3378.10 ± 0.5196.75 ± 0.41
Jointly learned82.87 ± 0.5478.14 ± 0.3296.73 ± 0.34
AD-DHGNN (ours)83.57 ± 0.1380.14 ± 0.1797.26 ± 0.11

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.