acceptodds
Under review as a conference paper at ICLR 2027

Beyond Pairwise Co-occurrence: Galois Closure Curvature for Higher-Order Relations

Abstract

Hypergraphs represent interactions among arbitrary sets of entities, yet hypergraph learning methods can often consume incidence primarily through pairwise co-occurrence, i.e., which nodes appear together. This reduction can obscure set-level dependencies: two hypergraphs may share every node degree and pairwise co-occurrence count while differing in which node sets necessarily imply others. We ask how the geometry of such higher-order dependencies can be characterized beyond observed connectivity. Our key observation is that Galois closure, which encodes implications among nodes induced by incidence, reveals this geometry through how closure states contract or expand under shared structural perturbations, following the local contraction principle underlying Ollivier–Ricci curvature. Building on this observation, we introduce Galois Closure Curvature (GCC), which quantifies the perturbation response of closure states, together with a curvature profile that retains contraction, expansion, and variability otherwise lost through cancellation in the scalar curvature. We show that this profile can distinguish hypergraphs with identical pairwise co-occurrence statistics even when their scalar GCC coincides, and characterize conditions under which the closure response vanishes on purely pairwise structures. By lifting the profile to node–hyperedge incidences, GCC provides a context-dependent structural representation of higher-order implication geometry. Across controlled implication-recovery tasks and established hypergraph benchmarks, GCC achieves the best mean rank among curvature descriptors on Hy-WL and is the most accurate on collections whose classes differ only in twin collections. The code is available at https://anonymous.4open.science/r/GCC-7A1C/.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Representation Learning, Metric Learning, Taxonomy Expansion

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_57qnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Galois Closure Curvature (GCC), a curvature for hypergraphs defined through the perturbation response of Galois closure states. Given two closure states, the method applies a shared perturbation and measures whether the states contract or expand, summarizing the response as a scalar curvature plus a three-component profile (positive response, negative response, variability). The authors prove that the profile can distinguish hypergraphs with identical degree and pairwise co-occurrence statistics, and even identical scalar curvature. Experiments cover hypergraph classification on Hy-WL, constructed twin collections designed to isolate set-level implications, and computation time on larger hypergraphs.

Critical strengths
  1. The core observation is real and cleanly demonstrated. Proposition 2 constructs two hypergraphs with identical B-transpose-B and identical scalar curvature, yet different profiles. This is a valid demonstration that scalar aggregation loses information, and it is not a contrived trick — the cancellation of contraction and expansion is a genuine phenomenon.

  2. The computational advantage over transport-based curvatures is concrete. GCC avoids optimal transport and computes responses by set operations. The timing results on the larger hypergraphs show a substantial speedup over ORCHID, HRCD, and ETR, which is a meaningful practical property.

Critical weaknesses
  1. The paper shows that the profile is discriminative, but not that it is useful beyond the diagnostic setting. The twin collections differ only in implication structure by construction, so high accuracy there mainly confirms that the profile detects the signal it was designed to capture. On the real Hy-WL collections the advantage is smaller, and on the two IMDB writer collections the method underperforms. The message-passing gains in the appendix (0.1–0.4 points) are within noise for most GNN benchmarks. A real task where implication structure is known to matter would strengthen the practical case.

  2. The relationship to existing curvatures is not fully clarified. GCC is compared against ORCHID, HFRC, and ETR, but the paper does not argue that the implication information it captures is absent from these descriptors. ORCHID's random-walk formulation operates on hyperedge transitions, a different object from closure perturbation, but the two may overlap in what they encode about intersection structure. A short discussion or ablation would clarify whether GCC's increment is orthogonal to these baselines.

  3. The scale of the validation is small relative to the claim. Hy-WL hypergraphs have roughly 10 to 35 nodes. The twin collections are synthetic. The larger hypergraphs are used only for timing, not for any downstream task. For a paper whose motivation is that higher-order implication structure matters in real data, there is no real large-scale task where that structure is shown to matter.

  4. The three profile components are not well interpreted. Contraction and expansion have clear readings, but variability does not: what does a high-variance response mean structurally, and when is it signal rather than noise? The paper reports an 8.6-point gain on RHG-10 from variability but does not explain why.

Questions for authors
  1. The twin collections isolate implication structure by construction. Can you provide a real hypergraph task, ideally at larger scale, where implication structure is known to matter and GCC outperforms existing descriptors? This would help establish the practical value of the profile.

  2. ORCHID's random-walk formulation involves transitions between hyperedges, which may relate to the intersection structure behind implications. Can you show empirically or analytically that GCC captures information not already present in ORCHID?

  3. What does a high-variability node pair or hyperedge correspond to structurally? Can you give a concrete example from one of the real collections where variability is the discriminating signal?

  4. The gains in Appendix E are small. Are these differences statistically significant under the same corrected resampled t-test used for the classification results? If not, what should a reader conclude about the practical utility of GCC as a structural encoding?

5 comments
AuthorsOpenReview ↗

Empirical Validation and Task-Dependent Effectiveness of GCC

We thank the reviewer for their feedback. We conducted analyses, direct comparisons, experiments and statistical tests to clarify GCC's higher-order expressivity, distinction from existing curvatures, structural interpretation, effectiveness, and limitations.

W1 & W3 & Q1. Real-World Evaluation and Structural Validation

Real-World Hyperedge Prediction. We evaluated GCC on hyperedge prediction across eight hypergraphs: four metabolic networks (node = metabolite, hyperedge = reaction; 2,746–8,346 nodes), where implications distinguish real from false reactions, and four citation hypergraphs from our timing study (node = paper, hyperedge = papers cited by one paper or written by one author; 2,708–19,717 nodes).

Table R3.1. Hyperedge prediction, AUC-PR (25-run mean). Rank: mean rank among twelve curvatures.

Recon3DHuman-GEMyeast-GEMRECON1Cora-CACiteseer-CCPubmed-CCCora-CCRank
Nodes5,8358,3462,7462,7642,7083,31219,7172,708–
Hyperedges8,57310,9203,8233,2439701,0047,5311,483–
Majority0.3330.3340.3340.3340.3680.3580.3340.337–
Size0.5340.5110.5180.5550.5990.5990.4700.597–
Degree0.5490.5310.5240.5610.5970.6070.4730.599–
HFRC0.5500.5380.5250.5590.6000.6060.4680.5998.00
CIHI0.5410.5160.5150.5540.5980.5900.4700.59510.75
ORCHID0.5450.5210.5190.5550.6000.6090.4700.5998.50
CE-FRC0.5740.5620.5240.5790.6170.6250.5310.6034.50
CE-ORC0.5840.5570.5830.6000.5960.6120.4810.6095.38
CE-AFRC30.6320.6250.5940.6150.5900.6060.4930.6093.88
ETR0.5870.5490.6200.6210.6360.6500.6380.6532.00
HRCD0.5480.5200.5200.5560.5990.5860.4710.5969.38
BE0.5380.5240.5180.5500.5960.5940.4690.6009.88
Res.0.5860.5610.5430.5660.6070.5950.4850.5926.25
GCC-κ0.5650.5190.5760.5930.6010.5900.4770.6007.25
GCC0.6290.5750.6080.6100.6500.6310.5130.6092.25

Following Yoon et al. (WWW 2020), all methods share 13 pairwise features; each row below Majority measures a descriptor's contribution. Majority, Size, and Degree are unranked controls (Table 1). Negatives replace one member of a held-out hyperedge. We conduct 25 runs with Appendix B's tests. Metabolic networks were selected using a separate split.

ETR ranks first among curvatures (2.00) but averages 189 s and 1,658 MB; GCC ranks second (2.25) at 0.45 s and 130 MB. GCC outperforms scalar GCC-κ, computed from the same closures and perturbations, on all eight hypergraphs.

Motivation for Twin Collection. Hypergraphs with identical degrees and pairwise co-occurrence can differ in higher-order implications. Testing requires (C1) whole-hypergraph labels and (C2) matched pairwise statistics.

Hy-WL satisfies C1, where GCC ranks first (2.80; next: 4.40), but not C2: hyperedge size alone achieves 99.9% on RHG-3. We enumerated 1,800 hypergraph pairs with identical B⊤BB^\top B and hyperedge sizes across five structural types (Appendix C). Descriptors based solely on matched statistics cannot distinguish these pairs, whereas GCC achieves 99.3%, versus 74.4% for the strongest curvature baseline.

GCC does not lead on IMDB writer collections, which have the lowest implication prevalence (7.9% and 7.0%, versus ≥27% elsewhere), consistent with its dependence on higher-order implications.

Q4. Significance of the Gains in Appendix E

Under Appendix B's corrected resampled t-test, neither GCC nor any of eight existing curvature encodings in Table 6 significantly improves over the backbone. This aligns with prior findings: structural encodings yield no significant gains on these datasets (Pellegrin et al., 2025), while curvature encodings benefit graph classification more than node classification (Fesser & Weber, 2024).

Curvature descriptors appear more effective for whole-hypergraph or node-set tasks (Tables 1 and R3.1) than node classification (Table 6). Table 6 predicts paper topics, where no encoding improves the backbone; in contrast, several curvatures including GCC outperform controls in Table R3.1. Message passing may capture much of the label-relevant structure in Table 6, leaving little benefit from encoding. The backbone is 7–16 points more accurate on nodes within hyperedges, with encoding changes limited to −0.07 to +0.08 points.

Revision. We will add the eight real-world hyperedge prediction results with runtime and memory comparisons, clarify the Twin collection's matched pairwise controls, and report corrected significance tests. We will also discuss GCC's task-dependent effectiveness and limitations.

AuthorsOpenReview ↗

Higher-Order Expressivity and Structural Distinctions from ORCHID

W2 & Q2. Relation to ORCHID.

ORCHID's probability measures are determined by statistics of node pairs, whereas GCC also uses statistics of node triples. GCC can therefore distinguish hypergraphs whose pair statistics coincide, and its gain over ORCHID on Hy-WL comes from this difference.

ORCHID's pair statistics. ORCHID's probability measures are one step of a random walk from node to node (Coupette et al., 2023, Eqs. 5–7). With cijc_{ij} the number of hyperedges that contain both ii and jj, they can be written for j≠ij\neq i as:

  • EN (equal nodes): μi(j)∝1[cij>0]\mu_i(j)\propto\mathbf{1}[c_{ij}>0];
  • WE (weighted edges): μi(j)∝cij\mu_i(j)\propto c_{ij};
  • EE (equal edges): μi(j)∝∑e∋i,j1/(∣e∣−1)\mu_i(j)\propto\sum_{e\ni i,j}1/(\lvert e\rvert-1).

All three weights are statistics of node pairs: whether ii and jj share a hyperedge (EN), how many hyperedges they share (WE) and how large these are (EE). A step of the walk may pass through a hyperedge, but μi(j)\mu_i(j) records only where the step ends and therefore sums over every hyperedge that contains ii and jj. Which further nodes such a hyperedge contains does not enter the measure. On intersection structure, the ORCHID paper itself notes that "hyperedge intersections can vary in cardinality, but this variation is not currently reflected in our probability measures" (Section 5).

GCC's triple statistics. GCC uses pair statistics as well, in its closure states and in the feasible perturbations of our Eq. (5). A node yy belongs to Ci=ϕ({i})C_i=\phi(\lbrace i\rbrace) iff ciy=ciic_{iy}=c_{ii}, and mm is a feasible perturbation for (i,j)(i,j) iff cim>0c_{im}>0 and cjm>0c_{jm}>0. Triples enter through the perturbed states of our Eq. (4): for m∉Cim\notin C_i, Tm(Ci)T_m(C_i) consists of the nodes that lie in every hyperedge containing ii and mm. These are the nodes implied by {i,m}\lbrace i,m\rbrace, and GCC compares them with the nodes implied by {j,m}\lbrace j,m\rbrace.

The following constructed example isolates this distinction under identical pair statistics:

H={abc,ade,bdf,cef,acf,aeg}H=\lbrace abc,ade,bdf,cef,acf,aeg\rbrace and H′={abd,ace,bcf,def,acf,aeg}H'=\lbrace abd,ace,bcf,def,acf,aeg\rbrace.

Both are 3-uniform and have identical pairwise co-occurrence counts; hence their EN, WE and EE measures and their ORCHID node-pair curvatures coincide. Their hyperedge-curvature distributions also coincide under exact W1W_1 with mean or max aggregation. Yet GCC assigns different profiles to abab: (0.10,0.60,0.55)(0.10,0.60,0.55) in HH and (0.20,0.30,0.58)(0.20,0.30,0.58) in H′H'. Perturbing both closure states by dd yields distinct sets (adeade, bdfbdf) in HH but the same set (abdabd) in H′H'. GCC detects this higher-order difference despite identical ORCHID measures.

On Hy-WL the GCC profile of 64% of the node pairs on average changes when only the co-occurrence graph is kept. The table below shows that GCC given only pair statistics performs at the level of ORCHID on average (71.4% vs. 71.0%) and gains 3.1 points with triple statistics.

DescriptorHy-WL mean accuracy (%)
ORCHID71.0
GCC (pair statistics only)71.4
GCC (w/ triple statistics)74.5

This suggests that the gain of GCC over ORCHID comes from the structure of the hyperedges beyond pairs, which ORCHID's measures do not use.

Revision. We will clarify the distinction between ORCHID's pairwise random-walk measures and GCC's triple-dependent closure perturbations, include the constructed example demonstrating their different expressivity under identical pair statistics, and add the Hy-WL ablation quantifying the contribution of higher-order information.

AuthorsOpenReview ↗

Structural Interpretation and Empirical Validation of Variability

W4 & Q3. Structural Meaning of Variability

High-variability pairs and hyperedges. A high-variability node pair is one whose common neighbours disagree: some separate uu and vv, others do not. A high-variability hyperedge is one whose node pairs are of this kind on average, since ce(ξ)c_e^{(\xi)} is the mean of ξuv\xi_{uv} over its node pairs (Eq. 8). This reading follows from the definition. For a node pair (u,v)(u,v), the shared perturbations mm are typically the two nodes themselves and their common neighbours. With d0=d△(Cu,Cv)d_0=d_\triangle(C_u,C_v) and dm=d△(Tm(Cu),Tm(Cv))d_m=d_\triangle(T_m(C_u),T_m(C_v)), the response is ρm=1−dm/d0\rho_m=1-d_m/d_0 and

ξuv2=Var(dm)d02=E[(dm−dm′)2]2d02,\xi_{uv}^2=\frac{\mathrm{Var}(d_m)}{d_0^2}=\frac{\mathbb{E}[(d_m-d_{m'})^2]}{2d_0^2},

where m,m′m,m' are independent feasible perturbations. Contraction and expansion say how much the perturbations pull the two closure states together or push them apart on average; variability says whether they agree: ξuv=0\xi_{uv}=0 iff every perturbation changes the distance by the same amount. A common neighbour mm can bind the pair (ρm=1\rho_m=1: adding mm sends both states to the same closed set, e.g. when uu, vv and mm lie together in a hyperedge), split it (ρm<0\rho_m<0: mm meets uu and vv in different hyperedges and sends the states into different sets; the larger these are, the more negative ρm\rho_m), or leave it unchanged. Geometrically, κ\kappa averages the response over perturbations as Ricci curvature averages sectional curvatures over directions, and ξ\xi measures the anisotropy: a pair can be flat in every direction, or saddle-like, with the same mean.

Signal versus noise. Variability carries structural information that κ+\kappa^+ and κ−\kappa^- do not determine. The same average pull and push can come from one strong perturbation or from several weak ones, and only ξ\xi tells these apart. For example, in IMDB-Wri-Genre two node pairs of one hypergraph both have κ+=κ−=4/7\kappa^+=\kappa^-=4/7, but one is pushed apart by two perturbations of equal strength (ρ=−2,−2\rho=-2,-2; ξ=1.31\xi=1.31) and the other by a single stronger one (ρ=−4\rho=-4; ξ=1.69\xi=1.69). This is frequent: in four of the six real collections, 53–96% of the node pairs with a non-zero profile share their exact (κ+,κ−)(\kappa^+,\kappa^-) with a pair of different ξ\xi (6–7% in the two IMDB writer collections). It is also signal in a checkable sense: if no perturbation pushes the states apart, all ρm∈[0,1]\rho_m\in[0,1] and hence ξ≤1/2\xi\le1/2, so ξ>1/2\xi>1/2 certifies a splitting neighbour.

Clarification on the 8.6-point gain on RHG-10. The 8.6 points compare the full profile (89.8%) with the scalar κ\kappa (81.2%), not a profile with and without ξ\xi. They arise because the scalar cannot separate two of the ten classes, whose mean responses coincide (AUC 0.50) while their spread differs; these two classes account for 8.4 of the 8.6 points. The ablation below shows the contribution of ξ\xi directly: adding it to [κ+][\kappa^+], [κ−][\kappa^-] and [κ+,κ−][\kappa^+,\kappa^-] raises the accuracy in every column. On the real collections the gain is clearest for IMDB-Wri-Genre (+4.0 points over κ+\kappa^+) and IMDB-Dir-Genre (+2.9 over κ−\kappa^-), both significant.

DescriptorRHG-10Hy-WL meanTwins K=3Twins K=5
[ξ]82.772.775.796.1
[κ⁺]78.572.482.596.6
[κ⁺, ξ]89.874.386.899.0
[κ⁻]80.571.872.893.0
[κ⁻, ξ]86.773.480.497.9
[κ⁺, κ⁻]88.674.386.598.2
[κ⁺, κ⁻, ξ] (GCC)89.874.588.699.3

Revision. We will expand Section 4.2 to explain variability as disagreement among shared perturbations, including its geometric interpretation and structural implications. We will also add the component-wise ablation results and clarify that the 8.6-point gain on RHG-10 reflects the full curvature profile rather than variability alone.

Reviewer_ncV1OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

In response to the issue that statistics on pairwise co-occurrence in hypergraphs fail to capture higher-order implications, this paper proposes the Galois Closure Curvature (GCC). Specifically, it first applies the same node perturbation to two Galois-closed states, then measures whether the graph contracts or expands based on changes in distance before and after the perturbation, and retains the corresponding positive and negative fluctuations as three-dimensional features to avoid the loss of information that would result from simply taking the mean. The authors provide a theoretical proof that such features can distinguish between hypergraphs that share identical node degrees, pairwise co-occurrence statistics, and even scalar GCC values, and aggregate them into structural descriptors suitable for hypergraph classification. Extensive experiments have demonstrated their effectiveness.

Critical strengths

(1) The theoretical framework is comprehensive and well-founded, with a clear and effective rationale. Proposition 2 clearly demonstrates the information that may be overlooked by pairwise co-occurrence, single-node closure and average perturbation response, respectively. This example effectively links the problem definition in the paper with the profile design, making the purpose of the method easy to understand. Multiple experiments demonstrate the effectiveness of GCC, which achieved the best average ranking on Hy-WL whilst incurring significantly lower computational costs than transmission-based curvature methods.

(2)An open-source code, whereby the author provides an accessible code repository, ensuring a high degree of reproducibility.

Critical weaknesses

(1) As the paper does not include ablation experiments, I am unable to determine which modules are responsible for the gains achieved by GCC. It would be advisable to supplement the paper with detailed ablation experiments on its key modules: ‘Galois Closure on Hypergraphs’, ‘Closure Perturbation’, ‘Curvature Perturbation-Aware Curvature Profiles’, and ‘Incidence-level lifting’.

(2) The signals in the twin collections originate from artificially introduced precise implications. However, as the Galois closure relies on strict hyperedge intersections, the addition of a single hyperedge that violates an implication may alter the closure. Consequently, the ‘state perturbation’ within the method is not equivalent to robustness against noise in the input data. The current experiments have not yet demonstrated to what extent GCC’s advantages are retained when some implications are broken, or associations are missing.

Questions for authors

(1) Under the same aggregation and SVM protocols, to what extent does GCC outperform simple descriptors such as binary closure increments? Does this explain its primary advantage over twin collections?

(2) If a small number of hyper-edges that violate the implanted implications were injected into existing twin collections, would this result in any changes, and how would GCC’s accuracy and relative advantage be affected?

5 comments
AuthorsOpenReview ↗

Component-wise Ablation and the Role of Incidence-Level Lifting

We thank the reviewer for the constructive feedback. We provide component-wise ablations to clarify the contribution of each GCC component and the role of incidence-level lifting.

W1. Component-wise Ablation

We evaluate four ablations under the same data splits, aggregation, and SVM evaluation protocol, replacing one component at a time while retaining the remaining pipeline:

  • w/o Galois Closure: Computes closure on a hypergraph reconstructed from pairwise co-occurrence rather than the original incidence structure. This tests whether the original higher-order incidence information is necessary.
  • w/o Closure Perturbation: Replaces perturbation-induced responses with static binary closure increments, without node toggling. This tests the contribution of perturbation-based structural responses beyond static closure information.
  • w/o Perturbation-Aware Profiles: Retains only scalar curvature κG\kappa_G (GCC-κ\kappa), discarding the separate contraction, expansion, and response variability components. This tests the value of preserving the signed response profile.
  • w/o Incidence-level Lifting: Replaces hyperedge-conditioned incidence profiles with context-free node-level profiles. This tests whether the organization of pairwise responses within hyperedges contributes beyond their global aggregation.
MethodHy-WL MeanTwin Mean
w/o Galois Closure71.460.4
w/o Closure Perturbation70.588.4
w/o Perturbation-Aware Profiles72.681.3
w/o Incidence-level Lifting74.578.1
GCC (Full)74.594.0

All four ablations reduce Twin accuracy. Removing Galois closure produces the largest reduction (94.0% to 60.4%), followed by removing incidence-level lifting (78.1%), perturbation-aware profiles (81.3%), and closure perturbation (88.4%). On Hy-WL, removing Galois closure, closure perturbation, or perturbation-aware profiles also reduces mean accuracy. Full GCC achieves the highest Twin mean and ties for the highest Hy-WL mean.

The different behavior of incidence-level lifting is particularly informative. Its removal leaves the Hy-WL mean unchanged (74.5%) but reduces Twin accuracy by 15.9 percentage points. This suggests that hyperedge-conditioned aggregation is especially useful for distinguishing the controlled higher-order structures in Twin Collection, although its benefit is not uniform across datasets.

The structural role of lifting follows from Eqs. (8)–(9). For a hyperedge ee with ∣e∣≥2|e|\geq2, the symmetric pairwise profiles guvg_{uv} are aggregated as

\qquad c_e=\frac{1}{\binom{|e|}{2}}\sum_{\{u,v\}\subseteq e}g_{uv}.$$ The resulting hypergraph descriptor is the empirical distribution of hyperedge profiles, $$\mu_H=\frac{1}{|E_{\geq2}|}\sum_{e\in E_{\geq2}}\delta_{c_e},$$ where $E_{\geq2}$ denotes the hyperedges containing at least two nodes. GCC compares these distributions through the component-wise Wasserstein distance in Eq. (9). Importantly, $c_e$ is an average of existing pairwise profiles, not a newly computed higher-order curvature response. The contribution of lifting is therefore representational: it preserves how pairwise responses are organized across individual hyperedges. Two hypergraphs may have similar global pairwise response statistics but different distributions of hyperedge-level profiles, which a global aggregation can obscure. This distinction also explains why lifting can improve Twin discrimination without necessarily improving the Hy-WL mean. Twin Collection explicitly controls lower-order statistics, making the organization of higher-order incidence information relevant to discrimination. In contrast, Hy-WL collections may contain predictive information in simpler statistics, so additional hyperedge-conditioned structure need not improve every collection. Overall, the ablations support complementary roles for the four components: original Galois closure retains incidence-dependent implications; closure perturbation measures structural responses; the signed profile avoids reducing these responses to a scalar; and incidence-level lifting preserves their hyperedge-conditioned distribution. These findings support the design choices without establishing that every component is independently necessary for every dataset. **Revision.** We will add the precise ablation definitions, results, and the mathematical interpretation of incidence-level lifting to the revised manuscript.
AuthorsOpenReview ↗

Structural Robustness and Discriminative Value Beyond Static Closure

W2 & Q2. Sensitivity to Hyperedge Perturbations

We evaluate the sensitivity of strict Galois closure using implication-violating insertions (1–2 hyperedges) and random deletions (10–20%) on Twin Collection. The same structural modifications are applied across methods (accuracy, %).

MethodK=5 Clean+1+2Del.10%Del.20%K=3 Clean+1+2Del.10%Del.20%
ETR74.471.469.468.766.065.263.562.060.659.1
GCC-κ\kappa92.386.679.986.281.770.265.862.767.665.4
GCC99.397.795.096.292.188.685.281.982.177.6

GCC achieves the highest accuracy in every tested condition, outperforming ETR and scalar GCC-κ\kappa. Although its accuracy decreases as perturbations intensify, the results show that useful discriminative information remains under the examined structural changes.

The closure response can be characterized exactly. For an inserted hyperedge ee,

\begin{cases} \phi_H(X)\cap e, & X\subseteq e,\\ \phi_H(X), & X\nsubseteq e. \end{cases}$$ Thus, insertion changes closure only when $X\subseteq e$ and $\phi_H(X)\nsubseteq e$. Conversely, deleting $e$ can affect closure only when $X\subseteq e$ and $e$ contributes to the intersection defining $\phi_H(X)$. Insertion can remove implied nodes from a closure, whereas deletion can only preserve or enlarge it, provided the closure query remains defined. This establishes that strict closure is sensitive but not uniformly affected by hyperedge modifications. Some closure states remain unchanged, while others may change substantially; the set of feasible perturbations may also change. Accordingly, we interpret the results as empirical robustness under the tested perturbation protocols, not invariance to arbitrary structural changes. ### Q1. Comparison with Simple Closure Descriptors To test whether GCC's gains arise merely from using closure information, we compare four static descriptors under the same downstream aggregation and SVM evaluation protocol. Let $C_u=\phi(\{u\})$, $C_v=\phi(\{v\})$, and $C_{uv}=\phi(\{u,v\})$: - **Binary increment (node pair):** $\mathbf{1}[C_{uv}\supsetneq\{u,v\}]$, indicating any additional implied node. - **Binary increment (closure states):** $\mathbf{1}[C_{uv}\supsetneq C_u\cup C_v]$, detecting implications that arise only jointly. - **Closure-size increment:** $|C_{uv}|-|C_u\cup C_v|$, counting these joint-only implications. - **Closure overlap:** $|C_u\cap C_v|/|C_u\cup C_v|$, measuring similarity between singleton closures rather than joint implications. | Method | K=3 | K=5 | |:--|--:|--:| | ETR | 65.2 | 74.4 | | Binary increment (node pair) | 85.9 | 95.2 | | Binary increment (closure states) | 83.1 | 93.6 | | Closure-size increment | 82.8 | 94.0 | | Closure overlap | 50.4 | 53.0 | | **GCC** | **88.6** | **99.3** | The strong binary-increment results show that closure information itself is discriminative. GCC nevertheless exceeds the best static descriptor by 2.7 percentage points at K=3 and 4.1 points at K=5. Thus, the improvement cannot be explained solely by the presence of closure-based features, although the comparison does not isolate every individual mechanism. The distinction can be expressed geometrically. Let $D=B\triangle C$ denote the disagreement between two closure states, and $D'=T_m(B)\triangle T_m(C)$ their disagreement after a matched perturbation. Then $$d_\triangle(B,C)-d_\triangle(T_m(B),T_m(C)) =|D\setminus D'|-|D'\setminus D|.$$ The first term counts disagreements resolved by the perturbation; the second counts newly introduced disagreements. A scalar net change can conceal simultaneous contraction and expansion, while GCC's signed profile retains their separate contributions and response variability. Consequently, static descriptors summarize closure implications, whereas GCC characterizes how differences between closure states respond to perturbations. This provides a geometric interpretation of the observed improvements without claiming that richer representations necessarily outperform static descriptors on every task. **Revision.** We will incorporate the perturbation settings, closure descriptor definitions, and corresponding structural and geometric analyses into the revised manuscript.
AuthorsOpenReview ↗

Supplementary Results: Full Ablation and Structural Robustness Evaluations

For reference, we provide the full experimental results supporting our responses to W1, W2, and Q2, including dataset-level component-wise ablations and structural robustness evaluations across all baselines (accuracy, %).

Supplementary Table 1. Component-wise Ablation Results

The table reports results across ten Hy-WL collections and Twin Collection, complementing the mean accuracies presented in W1.

MethodRHG-TRHG-PRHG-3RHG-10IDForIDGenIWForIWGenSteamTwittTwin K=3Twin K=5
w/o Galois Closure99.295.798.789.962.473.446.339.048.660.454.666.2
w/o Closure Perturbation100.096.095.879.463.172.444.233.762.258.483.193.6
w/o Perturbation-Aware Profiles100.096.095.481.266.777.243.443.362.660.270.292.3
w/o Incidence-level Lifting100.095.899.589.467.377.045.446.163.760.672.284.0
GCC (Full)100.095.899.589.868.377.545.243.164.560.888.699.3

Full GCC achieves the highest mean accuracy on Twin Collection. On Hy-WL, the effects of individual components vary across collections, consistent with the dataset-dependent contributions discussed in W1.

Supplementary Table 2. Hyperedge Perturbation Results

The table reports results under implication-violating hyperedge insertions (+1, +2) and random deletions (10%, 20%) on Twin Collection.

MethodK=5 Clean+1+2Del.10%Del.20%K=3 Clean+1+2Del.10%Del.20%
HFRC56.054.352.754.653.552.952.051.751.250.7
CIHI55.955.555.254.455.752.752.253.551.249.8
ORCHID54.353.552.853.353.451.550.951.951.552.6
CE-FRC51.851.150.250.450.350.450.550.550.350.0
CE-ORC54.353.452.753.353.451.550.951.851.552.6
CE-AFRC353.653.052.253.453.452.051.652.051.752.0
ETR74.471.469.468.766.065.263.562.060.659.1
HRCD71.568.266.967.464.464.262.161.161.359.3
BE58.256.754.856.953.654.353.553.653.051.5
Res.51.051.250.951.251.150.350.949.850.049.2
GCC-κ\kappa92.386.679.986.281.770.265.862.767.665.4
GCC99.397.795.096.292.188.685.281.982.177.6

GCC achieves the highest accuracy across all ten clean and perturbed conditions, extending the comparisons with ETR and GCC-κ\kappa in W2/Q2 to the full set of evaluated baselines.

Reviewer_z9C9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper aims to characterize set-level dependencies in hypergraphs that are not determined by node degrees or pairwise co-occurrence counts. It uses Galois closure to express which nodes are necessarily implied by a set of observed nodes. The core idea is to compare closure states under the same feasible node toggle and quantify the resulting contraction or expansion of their symmetric-difference distance. The proposed Galois Closure Curvature (GCC) averages these responses, while a three-component profile separately retains contraction, expansion, and variability to avoid cancellation. The profiles are lifted to incidences and aggregated into hypergraph descriptors for distribution-based comparison. Experiments report the best mean rank among the tested curvature descriptors on Hy-WL and improved recovery of planted implications.

Critical strengths
  1. The Introduction identifies a meaningful distinction between pairwise co-occurrence and set-level implication. Proposition 2 provides a concrete, verifiable example in which degrees, pairwise counts, and singleton closures coincide while the proposed profiles differ. This gives the motivation a precise mathematical basis and makes the connection between formal concept analysis and hypergraph representations worth investigating.
  2. Section 4.2 separates contraction, expansion, and dispersion instead of relying solely on their signed mean. Proposition 2 demonstrates that cancellation can hide nonzero responses, so retaining this information is a sensible design choice. The construction uses explicit set operations, and the incidence-level lifting provides a concrete route for using these responses as structural features.
Critical weaknesses
  1. The geometric interpretation of the prescribed coupling is insufficiently justified. Equation (6) uses a synchronous coupling over a feasible set that depends on both closure states. Thus, even the perturbation distribution associated with one state can change with its comparison partner. This differs from optimal transport between fixed statewise measures. The distinction is substantive: in Proposition 2's first hypergraph, the perturbed distributions of the singleton states for a and b are identical, assigning probability one-third to the empty set and two-thirds to the full vertex set. Their Wasserstein distance is therefore zero, giving Ollivier curvature 1 for these measures, whereas GCC is 0. This does not invalidate GCC as a response statistic, and the authors explicitly replace transport. However, the paper needs to explain what geometric principle justifies this particular coupling and what the curvature interpretation contributes beyond a closure-sensitivity statistic. Proposition 1's sign and range properties do not resolve this central methodological question.
  2. Singleton-seeded, single-toggle probing has a systematic blind spot for higher-arity implications. Sections 3.2 and 4.1 initialize each node with its singleton closure. When these closures are singletons, one toggle only examines empty, singleton, or two-node sets. If every pair is already closed, all responses can vanish despite nontrivial implications with larger antecedents. For a concrete example, add all 15 two-node hyperedges on the same six vertices to both type-5 motifs in Appendix C, Table 3. Their degrees, pairwise counts, and per-node hyperedge-size counts remain matched, and every GCC profile becomes zero. Nevertheless, the supported implication that {a,b,c} implies d holds only in the second augmented motif: the triple's closure is {a,b,c,d} there, versus {a,b,c} in the first. This does not contradict Proposition 2's existential separation result or require a complete invariant. It exposes an entire class of higher-order dependencies whose detectability is not characterized. That limitation is central to the paper's motivation and needs explicit theoretical treatment.
Questions for authors

see Weaknesses

3 comments
AuthorsOpenReview ↗

Clarifying the Geometric Foundation and Higher-Order Detectability of GCC

We thank the reviewer for their thoughtful and constructive feedback, particularly for recognizing the contributions of Proposition 2 and the signed curvature profiles. We hope the following responses help clarify our approach and address the concerns regarding geometric interpretation and higher-order detectability.

W1. Geometric Justification

Synchronous coupling. The distinction is intentional: GCC preserves the correspondence between structural interventions, whereas optimal transport permits outcomes from different interventions to be matched. This preserves intervention-specific contraction and expansion, consistent with established coupling-based analyses of metric contraction (Bubley & Dyer, 1997; Eberle et al., 2019).

For nonempty S=S(B,C)S=S(B,C) and distinct states B,CB,C, let μB∣C\mu_{B\mid C} and μC∣B\mu_{C\mid B} denote the distributions of Tm(B)T_m(B) and Tm(C)T_m(C) for uniform m∈Sm\in S. The synchronous coupling is valid, hence

\leq 1-\frac{W_1(\mu_{B\mid C},\mu_{C\mid B})} {d_\triangle(B,C)}.$$ Thus, GCC lower-bounds a *pair-conditioned* Wasserstein contraction coefficient, rather than representing conventional Ollivier–Ricci (OR) curvature (Ollivier, 2009). **Pair-dependent feasibility.** The common feasible set $S(B,C)=F(B)\cap F(C)$ retains interventions supported for both states. This avoids artificial closure changes arising from the convention $\phi(X)=V$ for unsupported sets ($\operatorname{supp}(X)=0$). Accordingly, pair dependence represents explicit conditioning on mutually supported interventions, rather than a fixed statewise Markov kernel. **Curvature over individual sensitivity.** As the reviewer observes, the identical post-perturbation marginals in Proposition 2 yield OR curvature $1$, whereas matched responses $(-1/2,-1/2,1)$ yield scalar GCC $0$ and profile $(1/3,1/3,1/\sqrt{2})$. This difference illustrates the role of intervention correspondence: optimal transport may rematch outcomes, while GCC preserves their distinct contraction and expansion effects. Moreover, GCC measures *between-state* distance deformation rather than individual state sensitivity. For $D=B\triangle C$ and $D'=T_m(B)\triangle T_m(C)$, $$d_\triangle(B,C)-d_\triangle(T_m(B),T_m(C)) =|D\setminus D'|-|D'\setminus D|$$ This connects GCC to pairwise metric contraction while distinguishing it from intrinsic OR curvature. ### W2. Higher-Order Detectability **Scope and vanishing-response condition.** The reviewer's counterexample identifies a meaningful boundary of single-toggle probing, which can be characterized through closure properties. Proposition 2 establishes *existential separation* beyond degree and pairwise co-occurrence, rather than complete higher-order detectability. More precisely, suppose $\phi(\varnothing)=\varnothing$, $\phi(\{v\})=\{v\}$ for every vertex, and $\phi(\{u,v\})=\{u,v\}$ for every supported pair. Then every feasible single-toggle on distinct singleton states preserves their distance: $$d_\triangle(T_m(C_u),T_m(C_v)) =d_\triangle(C_u,C_v)=2, \qquad \rho_m(C_u,C_v)=0.$$ Toggling either initial vertex yields an empty set and a closed pair; other toggles yield two overlapping closed pairs. Thus, higher-arity implications may coexist with vanishing GCC profiles, as in the reviewer's augmented type-5 motifs. We will extend Appendix A.5 accordingly. **Probing-depth characterization.** Importantly, this limitation depends on the probing depth rather than on the closure representation itself. With trivial singleton closures, an $r$-toggle operation $$T_M(C_v)=\phi(C_v\triangle M), \qquad |M|\leq r,$$ queries antecedents of size at most $r+1$. Conversely, any $k$-node antecedent $X$ can be directly queried using $k-1$ toggles by choosing $v\in X$ and $M=X\setminus\{v\}$. This characterizes accessible antecedent sizes without implying universal discrimination. **Constructive separation.** For the reviewer's augmented type-5 motifs, a two-toggle $M=\{b,c\}$ yields $\rho_M(C_a,C_d)=-1/2$ versus $1$. Uniform averaging over feasible perturbations with $|M|=2$ also produces distinct profiles for $(a,b)$: $$g_{ab}^{M_0,(2)} =(0,1/6,\sqrt{5}/6), \qquad g_{ab}^{M_1,(2)} =(1/6,1/6,1/2).$$ The resulting separation illustrates how increasing probing depth can recover structural distinctions missed by single-toggle probing. Together, the vanishing condition and probing-depth characterization identify the source of the limitation and a principled direction for extending the accessible structural information. **Revisions.** We will clarify pair-conditioned contraction versus OR curvature (Secs. 3.3–4.1), add the coupling bound, vanishing-response condition, and probing-depth results with proofs and calculations to Appendix A, and explicitly state the single-toggle expressivity boundary (Sec. 4.2). We will also refine the exposition throughout the manuscript to provide clearer intuition alongside the mathematical formulations and improve overall readability.
AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer z9C9,

Thank you again for your constructive feedback. We have uploaded the revised manuscript, which incorporates the changes described in our response (marked in red).

  • W1. Sections 3.3 and 4.1 and Appendix A.2 (Proposition 5) clarify the geometric interpretation of the synchronous coupling.
  • W2. Section 4.2 and Appendices A.7 and A.8 (Propositions 10 and 11) state the limit of single-toggle probing and characterize the effect of probing depth.

We would appreciate it if you could let us know whether our response and the revision address your concerns. We are happy to provide any further clarification.

Best regards, The Authors

The authors’ response has addressed my concerns. I will raise my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their constructive feedback. We have uploaded a revised manuscript, with changes marked in red. The main revisions are summarized below.

  • Geometric interpretation (Reviewer z9C9). We clarified the synchronous coupling and the pair-dependent feasible set (Sections 3.3 and 4.1) and added a bound that relates GCC to a pair-conditioned Wasserstein contraction coefficient (Appendix A.2).
  • Single-toggle probing (Reviewer z9C9). We stated its limit (Section 4.2) and added the vanishing-response condition and a probing-depth characterization (Appendices A.7 and A.8).
  • Ablation study (Reviewers ncV1 and 57qn). We added component-wise ablations (Section 5.5, Table 2), with the definitions of the variants, all subsets of the profile components and static closure descriptors in Appendix D.
  • Twin collections (Reviewers ncV1 and 57qn). We clarified the matched pairwise controls and added results under hyperedge insertions and deletions (Appendix C, Table 6).
  • Relation to ORCHID (Reviewer 57qn). We added a comparison of the statistics used by ORCHID and GCC, with a constructed example (Appendix A.9).
  • Additional tasks (Reviewer 57qn). We added hyperedge prediction on eight real hypergraphs with runtime and memory (Appendix F), and significance tests and a discussion of task dependence for node classification (Appendix G).
  • Exposition. We expanded the explanation of the variability component (Section 4.2) and of incidence-level lifting (Section 4.3).

We thank the reviewers again and are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.