acceptodds
Under review as a conference paper at ICLR 2027

Operator-Relative Energy: A Unified Framework for Understanding and Mitigating Over-smoothing in Hypergraph Neural Networks

Abstract

Repeated message passing in hypergraph neural networks (HGNNs) can suppress distinctions between node representations, a phenomenon known as over-smoothing that limits the benefits of deeper architectures. The resulting form of representational collapse depends on the underlying propagation rule. However, classical measures of over-smoothing, such as the Dirichlet energy, can fail to detect collapse when applied across different propagation rules. To address this limitation, we introduce an operator-relative energy that measures representation variation relative to the collapse behavior of a given propagation rule, while recovering classical Dirichlet energies as special cases. Using the proposed energy, we show that representation variation can decay exponentially with network depth across a broad class of hypergraph propagation operators, leading to over-smoothing, whereas initial residuals can prevent complete collapse by preserving a nonzero level of variation. Motivated by these theoretical results, we incorporate initial residuals and identity mappings (II) into five representative HGNN architectures to mitigate over-smoothing. Controlled propagation experiments on six real-world datasets demonstrate that the classical Dirichlet energy can fail to detect over-smoothing, while our operator-relative energy correctly detects representation collapse and confirms that initial residuals prevent the energy from vanishing. Node-classification experiments further show that the proposed II variants consistently improve predictive accuracy over their corresponding baseline architectures and sustain their performance as network depth increases.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hypergraph neural networks, Over-smoothing, Energy

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_bPtwOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies over-smoothing in hypergraph neural networks, where repeated message passing gradually removes distinctions between node representations and can make deeper models ineffective. Its central observation is that different propagation rules collapse representations in different ways, so standard Dirichlet-energy measures can incorrectly indicate that meaningful variation remains even after collapse has effectively occurred. To address this, the paper proposes an operator-relative energy whose notion of collapse is matched to the specific propagation rule being analyzed, giving a common framework across different hypergraph architectures.

The main theoretical contribution is a unified analysis showing when representation variation decreases with depth for a broad class of hypergraph propagation operators, while also recovering familiar Dirichlet energies as special cases of the proposed framework. The paper then uses this perspective to analyze initial residual connections and identity mappings, showing that these mechanisms can preserve a nonzero amount of representation variation instead of allowing complete collapse. The practical contribution is to incorporate these mechanisms into five representative hypergraph neural network architectures, thereby connecting the theoretical analysis directly to a concrete architectural strategy for mitigating over-smoothing.

Critical strengths
  1. The paper provides a principled framework for analysing over-smoothing across different hypergraph propagation rules. Rather than assuming that all propagation operators collapse node representations in the same way, the proposed operator-relative energy adapts the notion of representation variation to the specific collapse behavior of the operator, while recovering familiar energy measures as special cases. This is consequential because it turns an architecture-dependent phenomenon into a common analytical framework that can be applied across multiple model families.
  2. The theoretical analysis is supported by coherent experimental validation. The controlled experiments directly test whether the proposed energy detects collapse where conventional measures fail, while the classification experiments separately examine whether the theoretically motivated mitigation strategy improves robustness as network depth increases.
Critical weaknesses
  1. The conceptual novelty is narrower than the paper's broad framing may initially suggest, because operator-dependent limiting subspaces and rank collapse have already been established in prior work [1, 2]. Limitations of conventional Dirichlet-energy-style diagnostics for characterising over-smoothing have also been demonstrated [3]. Initial residuals and identity mappings were introduced by GCNII and subsequently incorporated into hypergraph neural networks through the UniGNN framework [4, 5]. More recent theory has further characterised how residual connections can prevent excessive smoothing and determine the limiting representation subspace [6]. Consequently, the main novelty appears to lie in the specific operator-relative energy construction and its unified hypergraph analysis, rather than in the underlying observations about operator-dependent collapse or residual-based mitigation.
  2. The empirical evidence is less conclusive than the headline best-of-depth results suggest because the best depth is selected using test performance and the reported per-comparison significance tests do not account for multiple comparisons.This issue is more consequential because recent published work such as FrameHGNN already addresses deep-HGNN over-smoothing using multi-frequency hypergraph filtering together with initial residual and identity-mapping mechanisms [7]. Moreover, DHG-Bench provides a substantially broader and standardised evaluation of deep hypergraph learning, so stronger comparison with this recent benchmark literature would help establish whether the empirical advantages extend beyond the selected baseline architectures [8].

References

[1] Graph Neural Networks Exponentially Lose Expressive Power for Node Classification, ICLR'20.

[2] Rank Collapse Causes Over-Smoothing and Over-Correlation in Graph Neural Networks, LoG'24.

[3] Are We Measuring Oversmoothing in Graph Neural Networks Correctly?, ICLR'26.

[4] Simple and Deep Graph Convolutional Networks, ICML'20.

[5] UniGNN: a Unified Framework for Graph and Hypergraph Neural Networks, IJCAI'21.

[6] Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs, ICLR'25.

[7] Deep Hypergraph Neural Networks with Tight Framelets, AAAI'25.

[8] DHG-Bench: A Comprehensive Benchmark for Deep Hypergraph Learning, ICLR'26.

Questions for authors
  1. Could the authors recompute the best-depth comparisons using depth selected exclusively by validation performance, followed by evaluation on untouched test results, and report multiplicity-adjusted inference across the architecture-dataset comparisons? The appendix explicitly states that best depth is currently chosen according to test mean and that the reported p-values are uncorrected.
  2. How do the proposed II variants compare under the same splits and depth protocol with a recent published deep-HGNN method designed explicitly for over-smoothing [7]? I would not require the proposed approach to outperform it, but such a comparison would help determine whether the empirical contribution is competitive with the current literature rather than only with unmodified shallow architectures.
2 comments

We thank the reviewer for these valuable comments. We have revised the manuscript with changes highlighted in brown and red. Below, we address each point concisely.

Novelty.

We thank reviewer for this careful read. Our contribution lies in the operator-relative energy construction and its unified hypergraph analysis, not in the observation of operator-dependent collapse [1, 2], the limitations of Dirichlet-type diagnostics [3], or the II mechanism itself [4, 5]. We have revised the abstract, introduction, and contribution list to reflect this. Beyond these prior observations, our framework makes three distinct theoretical contributions: (i) an operator-relative energy whose zero set exactly matches the operator-specific collapse subspace; (ii) a unified exponential decay analysis for non-symmetric two-stage hypergraph operators, including intermediate activations and learned channel transformations; and (iii) explicit deterministic lower bounds quantifying the variation preserved by II. In particular, unlike [6], which studies linearized graph networks with symmetric operators, our analysis extends to two-stage hypergraph operators and provides both asymptotic and finite-depth guarantees under explicit conditions.

1. Best-depth comparisons.

We thank the reviewer for this comment. The depth of each model is now selected by mean validation accuracy, and test accuracy is reported only at that depth. P-value is adjusted by Benjamini-Hochberg procedure controlling the false discovery rate at 0.05.

Overall, the direction of the comparison is unchanged. The II variant attains a higher mean test accuracy than its baseline in 27 of 30 pairs, and the three exceptions are each within one reported standard deviation. The strength of the evidence at the selected depth is lower than we originally reported, with 11 improvements remaining significant after correction. We have revised the results section accordingly, and the appendix now describes this protocol in place of the earlier caveat. The table below summarizes the changes.

Submitted versionRevised version
II variant better than baseline28 of 3027 of 30
Significant improvements20 of 3011 of 30
II variants selecting L∈{16,32,64}L\in\{16,32,64\}23 of 3014 of 30

2. FrameHGNN.

We thank the reviewer for this suggestion and have added a comparison with FrameHGNN [7] at matched depths. FrameHGNN uses the same initial residual and identity mapping as the II variants and adds a learned filter over multiple hypergraph frequencies.

Due to time limitations, some runs at 3232 and 6464 layers and those on 20News100 and Pokec are still in progress. The table below reports the 1919 dataset--depth combinations completed so far, and the remaining ones will be added to the revised manuscript.

These results suggest that the II variants are competitive with FrameHGNN. FrameHGNN has the higher mean in 10 of the 19 combinations and the best II variant in the other 9, with a largest gap of 0.890.89 points. No significant difference was found between FrameHGNN and the best II variant at any matched depth after multiplicity correction.

DatasetDepthHGNN-IIHNHN-IIHyperGCN-IIAllSetformer-IIUniGCNIIFrameHGNN
Cora2279.68 ±{\pm} 0.7878.20±1.7978.20 {\pm} 1.7979.35±1.3279.35 {\pm} 1.3274.84±2.1574.84 {\pm} 2.1581.08±2.21\boldsymbol{81.08 {\pm} 2.21}80.69±1.6280.69 {\pm} 1.62
4478.83±1.0078.83 {\pm} 1.0077.99±2.0077.99 {\pm} 2.0079.22±1.7279.22 {\pm} 1.7276.26±1.4176.26 {\pm} 1.4181.11±1.73\boldsymbol{81.11 {\pm} 1.73}80.28±1.6880.28 {\pm} 1.68
8879.73±1.0079.73 {\pm} 1.0077.74±1.8777.74 {\pm} 1.8780.27±1.7180.27 {\pm} 1.7177.53±1.1477.53 {\pm} 1.1480.16±1.5780.16 {\pm} 1.5780.56±2.09\boldsymbol{80.56 {\pm} 2.09}
161679.78±1.1879.78 {\pm} 1.1878.83±2.0278.83 {\pm} 2.0278.76±1.4478.76 {\pm} 1.4478.05±1.0678.05 {\pm} 1.0680.78±1.5080.78 {\pm} 1.5080.92±1.64\boldsymbol{80.92 {\pm} 1.64}
Citeseer2274.40±0.94\boldsymbol{74.40 {\pm} 0.94}73.47±1.2273.47 {\pm} 1.2271.81±1.1671.81 {\pm} 1.1671.88±1.5271.88 {\pm} 1.5274.18±0.9874.18 {\pm} 0.9873.97±1.4573.97 {\pm} 1.45
4473.56±1.5273.56 {\pm} 1.5272.26±0.9072.26 {\pm} 0.9073.33±1.3873.33 {\pm} 1.3873.08±1.1173.08 {\pm} 1.1173.79±1.1473.79 {\pm} 1.1474.15±1.21\boldsymbol{74.15 {\pm} 1.21}
8874.05±1.2974.05 {\pm} 1.2973.29±1.0673.29 {\pm} 1.0673.25±0.9773.25 {\pm} 0.9773.67±0.9873.67 {\pm} 0.9874.12±0.8974.12 {\pm} 0.8974.46±0.98\boldsymbol{74.46 {\pm} 0.98}
161673.96±1.0173.96 {\pm} 1.0173.00±0.9973.00 {\pm} 0.9972.42±1.4572.42 {\pm} 1.4573.49±1.2073.49 {\pm} 1.2074.29±0.75\boldsymbol{74.29 {\pm} 0.75}74.06±1.1874.06 {\pm} 1.18
323273.86±1.3673.86 {\pm} 1.3672.52±1.5872.52 {\pm} 1.5872.86±0.9672.86 {\pm} 0.9673.39±1.0773.39 {\pm} 1.0774.28±1.59\boldsymbol{74.28 {\pm} 1.59}74.05±1.1174.05 {\pm} 1.11
House-Committee2284.37±2.3884.37 {\pm} 2.3883.99±2.3683.99 {\pm} 2.3683.22±1.5883.22 {\pm} 1.5882.97±1.8582.97 {\pm} 1.8585.39±1.56\boldsymbol{85.39 {\pm} 1.56}84.71±1.7384.71 {\pm} 1.73
4483.50±2.2383.50 {\pm} 2.2385.02±2.1885.02 {\pm} 2.1884.06±2.1584.06 {\pm} 2.1584.46±2.1084.46 {\pm} 2.1085.33±2.00\boldsymbol{85.33 {\pm} 2.00}84.95±2.0284.95 {\pm} 2.02
8883.75±2.3483.75 {\pm} 2.3484.37±1.7784.37 {\pm} 1.7784.30±1.5084.30 {\pm} 1.5083.62±2.6483.62 {\pm} 2.6485.33±1.71\boldsymbol{85.33 {\pm} 1.71}84.95±1.9284.95 {\pm} 1.92
161684.92±2.2784.92 {\pm} 2.2783.19±2.7183.19 {\pm} 2.7184.09±1.5584.09 {\pm} 1.5584.30±1.9684.30 {\pm} 1.9685.45±2.1085.45 {\pm} 2.1085.60±1.21\boldsymbol{85.60 {\pm} 1.21}
323284.09±2.3584.09 {\pm} 2.3584.37±1.7484.37 {\pm} 1.7484.02±1.6784.02 {\pm} 1.6783.68±2.5283.68 {\pm} 2.5284.33±2.2484.33 {\pm} 2.2485.26±1.75\boldsymbol{85.26 {\pm} 1.75}
NTU20122288.91±0.9988.91 {\pm} 0.9986.70±0.7786.70 {\pm} 0.7786.40±1.2886.40 {\pm} 1.2885.65±2.2185.65 {\pm} 2.2188.89±1.3988.89 {\pm} 1.3989.07±1.06\boldsymbol{89.07 {\pm} 1.06}
4488.61±1.5488.61 {\pm} 1.5486.94±1.1486.94 {\pm} 1.1486.30±2.5786.30 {\pm} 2.5787.71±1.7887.71 {\pm} 1.7888.41±1.3688.41 {\pm} 1.3689.15±1.47\boldsymbol{89.15 {\pm} 1.47}
8889.52±1.3589.52 {\pm} 1.3587.02±1.3587.02 {\pm} 1.3587.99±1.4087.99 {\pm} 1.4088.03±1.0288.03 {\pm} 1.0289.46±1.5389.46 {\pm} 1.5389.58±1.33\boldsymbol{89.58 {\pm} 1.33}
161689.50±1.7989.50 {\pm} 1.7986.94±1.5986.94 {\pm} 1.5985.79±2.4285.79 {\pm} 2.4288.03±1.1488.03 {\pm} 1.1489.26±1.4889.26 {\pm} 1.4889.82±0.91\boldsymbol{89.82 {\pm} 0.91}
323289.64±1.4289.64 {\pm} 1.4284.75±1.7384.75 {\pm} 1.7378.81±1.8978.81 {\pm} 1.8986.30±2.2486.30 {\pm} 2.2489.74±1.53\boldsymbol{89.74 {\pm} 1.53}89.64±1.4089.64 {\pm} 1.40

Node classification accuracy (%) at matched network depths. Entries give the mean ±\pm standard deviation of test accuracy over ten seeds.

Reviewer_q3LeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the oversmoothing prblem in hypergraph neural networks, where authors argue that classical Dirichlet energies can have mismatched subspace for capturing the oversmoothing effect. The authors therefore propose operator-relative energy, justified it both theoretically and empirically, as well as proposing some mitigation methods.

Critical strengths
  1. There are relatively fewer papers studying oversmoothing in hypergraphs, and the authors address this gap by proposing a hyperparagraph-specific energy formulation for measuring oversmoothing in GNNs
  2. The experiments are broad and cover common datasets and HGNN architectures
Critical weaknesses

The novelty of the paper's contribution is limited on the following two fronts:

  1. Oversmoothing metrics with mismatched subspaces are known to fail as discussed in [1, 2, 3], and the core proposition of the energy formulation in Eq 3 and 4 just replaces the subspaces / eigenvectors in the classical Dirichlet formulation with the correct operator-specific ones.
  2. The theory on the energy decay bound largely adapts proof from the existing literature [4, 5].
  3. Although already cited in the related works, the proposed oversmoothing alleviation method is similar to [6].

[1] Roth, Andreas. "Simplifying the theory on over-smoothing", arXiv:2407.11876, 2024

[2] Bison, Anna, and Alessandro Sperduti. "Analysis of Dirichlet Energies as Over-smoothing Measures" arXiv:2512.09890, 2025

[3] Kaicheng Zhang, et. al. "Are we measuring oversmoothing in graph neural networks correctly?", ICLR, 2026

[4] Chen Cai and Yusu Wang. "A note on over-smoothing for graph neural networks", ICML GRL+ Workshop, 2019

[5] Chen, Guanzi, et al. "Preventing over-smoothing for hypergraph neural networks", arXiv:2203.17159, 2022.

[6] Scholkemper, Michael, et al. "Residual connections and normalization can provably prevent oversmoothing in gnns", ICLR, 2025.

Questions for authors

It would be helpful if the authors can pleaes clarity the concerns related to novelty and better distinguish the works from the existing ones as mentioned in the weaknesses

1 comment

We sincerely thank the reviewer for pointing out references [1--6], which provide a great opportunity to clarify the core distinctions between our work and prior studies. All changes in the revised manuscript are highlighted in green and red. In summary, our central contribution is a unified operator-relative energy framework that connects the diagnosis, theoretical characterization, and mitigation of over-smoothing across different hypergraph architectures. Below, we address each point concisely.

1. Subspace mismatch ([1--3]).

We fully agree that [1--3] have already identified subspace mismatch, and we do not claim this observation as our contribution.

However, their remedies take a different route from ours. [2] just suggested to pick the matching choice. It does not, however, provide a construction of the matching energy for a general operator, and it does not consider non-symmetric operators. [1] and [3] replace energies with rank-based measures, which do not identify where the representations collapse to and do not provide a per-layer decay rate through learned weights and activations. Furthermore, the rank couldn't tell over-smoothing from over-correlation. Rank one means X=uv⊤X = uv^\top for any uu. If uu separates the nodes well, rank reports collapse while nodes remain fully distinguishable. While our energy is zero only when uu is the collapse direction. And None of [1--3] treats hypergraph operators.

We instead think the energy approach can be kept if it is built from the operator's own eigenvectors. The definition in Eq. (3) is simple by design. Its zero set is exactly the operator's collapse subspace (Definition 4.1). It satisfies an explicit per-layer decay bound (Theorem 4.4). Under II, its limiting value is bounded below by α2,EP(X0)\alpha^2,\mathcal{E}_P(X^0) (Theorem 4.6). Thus, unlike existing diagnostics, our energy is not merely a measure of collapse, but also a tool for establishing theoretical guarantees on both its emergence and prevention.

2. Energy decay ([4, 5]).

We agree that the skeleton is the same, because our work and [4,5] all belong to the paradigm of spectral energy analysis. However, we have additional essential add-ons.

First, the analyses of [4,5] are strictly confined to the simplest single-stage symmetric GCN and HGNN operators, where a Euclidean orthogonal eigenbasis is available. Their proofs do not apply as they stand to non-symmetric two-stage HGNNs such as HNHN and UniGCN, which have no such eigenbasis. Our operator-relative energy and the shared-incidence factorization supply the missing piece. The weighted adjointness ΠB=A⊤ΠE\Pi B = A^\top \Pi_E makes PP self-adjoint in the Π\Pi-weighted inner product (Proposition 4.3), so the spectral argument goes through.

Second, the same adjointness gives a hyperedge energy and the half-step bounds EE(AX)≤λ2 EP(X)\mathcal{E}_E(AX) \le \lambda_2\,\mathcal{E}_P(X) and EP(BZ)≤λ2 EE(Z)\mathcal{E}_P(BZ) \le \lambda_2\,\mathcal{E}_E(Z), which have no counterpart in [4,5]. These stagewise bounds are essential for handling intermediate nonlinear activations and learned channel transformations, which cannot be addressed by directly applying the existing single-stage spectral arguments. With these, Theorem 4.4 extends the decay bound to all six architectures in Table 1 and recovers the HGNN result of [5] as a special case, which directly demonstrates the generalizability and mathematical unity of our proposed energy framework.

3. Similar alleviation ([6]).

The II mechanism is not ours or [6]'s. It comes from GCNII and its hypergraph adaptations Deep-HGCN and UniGCNII. Our contribution regarding II is a general, quantitative theory of how it preserves variation outside operator-specific collapse subspaces. In particular, our work differs from [6] in two aspects.

Setting. [6] studies linearized graph networks with a symmetric operator and states that non-symmetric operators are outside its scope. We treat two-stage shared-incidence operators, which are non-symmetric in general.

Guarantee. [6] shows that a collapse measure stays above some constant c>0c > 0, with high probability for Gaussian weights and deterministically under regularity conditions, and characterizes the reachable subspace. We give an explicit deterministic floor, EP(X⋆)≥α2EP(X(0))>0\mathcal{E}_P(X^\star) \ge \alpha^2 \mathcal{E}_P(X^{(0)}) > 0, with sharp constants in Proposition B.8 and a finite-depth version in Theorem A.1. Unlike a generic non-collapse guarantee, our result quantitatively characterizes the variation preserved outside each operator's collapse subspace, with an explicit dependence on the residual strength α\alpha and initial energy.

Reviewer_RX6zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an operator-relative energy, E_P, for hypergraph neural networks. This energy is zero exactly when node features lie in the propagation operator's collapse subspace. It recovers the standard unnormalized and normalized hypergraph Dirichlet energies as special cases (Remark 4.2). Under the stated assumptions, the authors prove exponential energy decay for two-stage layers with a shared-incidence factorization (Theorem 4.4). Under further assumptions, they show that initial residuals and identity mappings (II) preserve a limiting energy of at least α² times its initial value (Theorem 4.6). Controlled linear experiments illustrate how classical Dirichlet energies can miss operator-specific collapse. Classification experiments show that the evaluated II versions of five HGNN models remain stable as depth increases to 64 layers.

Critical strengths
  1. The motivation is clear. Different propagation operators can have different collapse subspaces, so an energy designed for one operator may fail to detect collapse under another. Section 3 gives a simple example. Theorem B.5 and Corollary B.6 formalize a related limitation: no single nontrivial seminorm can decay under every operator in the full shared-incidence family with spectral radius one.

  2. The theoretical analysis on the proposed energy to include several HGNNs as special cases. Some lower bounds are also reported after including II.

  3. Experimental results (Figure 2) show that the evaluated II versions are more robust to increasing depth than the corresponding baselines across the tested models and datasets.

Critical weaknesses
  1. The authors do not discuss the technical challenges of the proposed energy and the provided theory, along with the main ideas to overcome the challenges. The paper is hard to follow in the technical parts, but the technical ideas appear to be not involved.

  2. While there are several message-passing HGNNs placing under the umbrella, it is equally important to character why HGNN methods counting oversmoothing, such as ED-HNN and UniGCNII, would perform under the proposed energy.

  3. Including II into GNN/HNN is not new. UniGCNII has analyzed the adaptation of GCNII into hypergraphs. Therefore, it is not surprising after including II the oversmoothing can be alleviated.

Additional comments:

Some baseline and II versions differ in settings beyond II. Appendix C reports several additional differences:

(a) HNHN base enables the intermediate activation, while HNHN-II disables it.

(b) The HyperGCN-II search additionally considers lazy propagation, (I + P)/2. The manuscript does not state which selected configurations use this option.

(c) The UniGCNII optimizer fixes its learning rate at 0.01 and uses weight decay values of 0.01 and 0.0005 for its two parameter groups. The searched learning rate and weight decay values do not override these settings.

These differences make it harder to separate the contribution of II from other configuration changes.

Questions for authors

Please see the weaknesses above.

1 comment

We thank the reviewer for these comments, which have helped us clarify the technical contributions and streamline the presentation. All changes in the revised manuscript are highlighted in blue and red. Below we address each point.

1. Technical Challenges

In the revised manuscript, (i) the abstract and introduction now state each technical challenge and our solution, and (ii) Section 4 adds plain-language explanations that connect the results. In short, there are two major challenges.

The first challenge lies in constructing the energy. Existing remedies either just advise matching the energy to the model without providing an explicit construction, or against the energy by replacing with rank-based measures. We instead think the energy approach can be kept if it is built from the operator's own eigenvectors. Therefore, we address this challenge by introducing an operator-specific weighting and deriving an explicit operator-relative energy whose zero set matches the propagator's collapse subspace.

The second challenge lies in analyzing the decay under non-symmetric propagation. Existing decay proofs rely on a Euclidean orthogonal eigenbasis, which is available only for symmetric operators. Two-stage operators such as HNHN and UniGCN have none. We establish weighted adjointness, ΠB=A⊤ΠE\Pi B=A^\top\Pi_E, which makes PP self-adjoint in the Π\Pi-weighted inner product. This provides an orthogonal eigenbasis under the Π\Pi-weighted inner product, enabling us to extend the spectral analysis to the non-symmetric setting.

2. Existing Anti-over-smoothing Methods

UniGCNII. Its propagation update is a special case of Eq. (8), obtained by setting ME,l=IM_{E,l}=I and ϕE(Z)=Z\phi_E(Z)=Z. Our II analysis therefore applies directly under the stated assumptions. In particular, Theorem 4.6 gives EP(X⋆)≥α2EP(X(0))\mathcal{E}_P(X^\star)\ge\alpha^2\mathcal{E}_P(X^{(0)}), establishing a positive limiting energy whenever the initial energy is nonzero. This explains its anti-over-smoothing mechanism through the preservation of variation outside the UniGCN operator's collapse subspace.

ED-HNN. Its general form is not directly covered by our current theory. The key distinction lies in the hyperedge-to-node update. Our model aggregates hyperedge representations using fixed, nonnegative weights shared across feature channels. ED-HNN instead computes a nonlinear message for each receiving node before aggregation:

mv(l)=∑e∋vρ(xv(l),ze(l)),ze(l)=∑u∈eφ(xu(l)).m_v^{(l)}=\sum_{e\ni v}\rho \left(x_v^{(l)},z_e^{(l)}\right),\qquad z_e^{(l)}=\sum_{u\in e}\varphi \left(x_u^{(l)}\right).

Because ρ\rho jointly depends on the receiving node's representation and the hyperedge representation before aggregation, it generally cannot be absorbed into our fixed aggregation matrix BB or the node activation ϕV\phi_V applied after aggregation. The general ED-HNN update therefore requires additional structural assumptions or a separate analysis to obtain energy-decay or preservation guarantees.

We have revised the discussion to incorporate this as an interesting future work.

3. Novelty of II

We adopt II as an established mechanism introduced by prior works, and didn't claim II as our contribution. Our novelty lies in the unified theoretical framework: we introduce an operator-relative energy that captures architecture-specific representation collapse, establish conditions for exponential energy decay across different hypergraph propagation operators, and prove that II preserves a strictly positive asymptotic energy floor under explicit conditions. These results provide a theoretical explanation and quantitative guarantees for the effectiveness of II beyond specific architectures. We have revised the manuscript to clarify this point.

4. Additional Comments

We thank the reviewer for identifying these inconsistencies. We have revised the implementations and experimental settings so that each Base–II pair differs only in the initial residual and identity-mapping operations. In addition, the depth of each model is now selected by mean validation accuracy, and test accuracy is reported only at that depth. Significance is assessed a two-sided Welch tt-test, with Benjamini-Hochberg correction. Results are updated in the manuscript Figure 2 and Table 3.

Implementation changes are as follows:

  1. Both HNHN and HNHN-II now enable the intermediate activation.
  2. Both HyperGCN and HyperGCN-II use the fast mediator graph. We have removed the additional lazy-propagation option (I+P)/2(I+P)/2 from its search space.
  3. UniGCN and UniGCNII now use the same fixed learning rate and weight decay.

Overall, the direction of the comparison is unchanged, but the strength of the evidence at the selected depth is lower than we originally reported, with 11 improvements remaining significant after correction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.