acceptodds
Under review as a conference paper at ICLR 2027

Intervention-Verifiable Hierarchical Self-Explaining Hypergraph Nueral Networks

Abstract

Hypergraph neural networks effectively capture higher-order interactions, but identifying the structural basis of their predictions remains challenging. Post-hoc explainers attribute predictions to explanation subhypergraphs or hyperedge subsets after training without establishing whether the identified structure drives the output. To this end, we introduce Intervention-Verifiable Hierarchical Self-Explaining Hypergraph Neural Network (IV-SENN), a framework that enables direct verification of prediction changes under structural interventions on hypergraphs. IV-SENN represents higher-order hypergraph structure through h-motifs and captures recurring feature patterns using learnable feature prototypes. The additive concept-interaction head decomposes each class logit into linear, compositional, and quadratic interaction terms over the full concept encoding. With the trained head held fixed, IV-SENN exactly decomposes intervention-induced logit changes into structural and feature-linked contributions for counterfactual verification. Global explanations identify class-critical h-motifs from the learned structural weights of the additive concept-interaction head. Local explanations report signed structural and feature concept contributions to individual node predictions and ground influential structural concepts in concrete h-motif instances. Across four real-world datasets, IV-SENN remains competitive in node classification while attaining the highest Fid_KL and lowest SHD. On N-SYNTH, its estimated intervention effects achieve a Pearson correlation of 0.95 with the ground-truth changes in the predefined score.

Primary area:
interpretability and explainable AI
Keywords:
Hypergraph Learning, Self-Explaining Neural Networks, Intervention Consistency, Hierarchical Semantic Anchoring

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AqjEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

IV-SENN predicts node labels using normalized h-motif counts, learned feature prototypes, and a quadratic concept-interaction head. It provides exact concept-level logit decompositions and recomputes changes under structural edits with fixed model parameters. A separate convolutional branch supports local motif grounding. Experiments cover four real hypergraphs and a synthetic benchmark for motif recovery and intervention-effect comparison.

Critical strengths
  • Using h-motifs directly preserves higher-order structural semantics and connects concept contributions to concrete hyperedge patterns. This is a useful explanation interface for hypergraphs.
  • The evaluation includes retained/removed-subgraph fidelity, stability, planted-signal recovery, and null-intervention controls. The reported fidelity and stability improvements provide empirical support beyond classification accuracy.
Critical weaknesses
  • The local grounding rule can miss interaction-driven concepts. Equation 37 selects candidate motif types using only effective linear weights, before Equation 27 considers their full contributions. Important types with weak linear weights may therefore be excluded. It is unclear how often this affects the reported explanations.
  • The methodological and theoretical advances appear incremental. The compositional map is a linear reparameterization (Proposition 5(a)), and the convolutional branch does not enter prediction. Theorem 1 then formalizes exact differences of the resulting quadratic head. Table 3 supports the design empirically, but the conceptual advance beyond this transparent predictor remains modest.
  • N-SYNTH measures agreement with a predefined quadratic reference score that does not generate the training labels (Appendix C.1). This is a useful controlled check, but its interpretation is limited to that reference. I therefore place less weight on this result as evidence of broader intervention-effect recovery.
Questions for authors

Please see the weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to CW1-CW2

Thank you for identifying aspects of the manuscript that needed further clarification. We quantified how often candidate screening changes the selected explanation, clarified the link between motif-based prediction and hyperedge edits, and added a synthetic experiment with known probability responses. We address each of your concerns below.

CW1:Interaction-Driven Motifs Missed by Candidate Screening

Response: Thank you for pointing this out. We agree that a motif type with a small effective linear weight can still make a substantial contribution to the class prediction through interactions. Equation (37) uses only the magnitude of the effective linear weight to screen types, whereas the full contribution of a local type kk is ϕv,k(c)=sv,k(w~k,cstruct+Iv,k(c))\phi_{v,k}^{(c)} = s_{v,k}(\widetilde w_{k,c}^{\mathrm{struct}} + I_{v,k}^{(c)}), where Iv,k(c)I_{v,k}^{(c)} is the interaction term. Let MvM_v be the largest positive contribution among the retained local types, with zero as the baseline. An excluded type with sv,k>0s_{v,k} > 0 exceeds MvM_v whenever Iv,k(c)>Mv/sv,k−w~k,cstructI_{v,k}^{(c)} > M_v/s_{v,k} - \widetilde w_{k,c}^{\mathrm{struct}}. In the pooled audit of Cora and DBLP, 4.4% of evaluated nodes had an excluded local type whose full contribution exceeded that of every retained local type.

To measure how often screening changed the final explanation, we held the predictor and three-hyperedge budget fixed and compared the original selection with the selection obtained by admitting all local motif types to Eq. (27). The selected explanation changed for 59 of 1,894 nodes on Cora (3.12%) and 710 of 36,433 nodes on DBLP (1.95%). An omitted type can rank higher by full contribution and still receive a lower grounding score once role evidence and instance matching are included. Across both coauthorship datasets, candidate screening preserved the selected explanation for the vast majority of evaluated nodes, showing that the rule can narrow the candidate set while keeping the final instance selection largely stable.

We added the screening condition and the audit to Appendix D.1 (p. 19, lines 1008–1016). Your comment also motivates future work on how to incorporate interaction contributions into screening while limiting the number of candidates.

CW2:Methodological and Theoretical Contributions

Response: Thank you for raising this point about the methodological and theoretical contributions. The original manuscript devoted more space to the quadratic head and could have made clearer how h-motif encoding, instance selection, and intervention analysis work together to explain hypergraph predictions. As a result, our advances in explanation at multiple levels and in hyperedge intervention analysis were less apparent.

Our main contribution is to connect the higher-order structural concepts used by the predictor to specific hyperedges in the original hypergraph and test the model’s response to edits of those hyperedges. To our knowledge, IV-SENN is the first self-explaining hypergraph model to use h-motifs as explicit structural concepts in its prediction path. An h-motif describes an overlap pattern among three connected hyperedges, giving prediction and explanation a common structural unit.

On the methodological side, we propose a rule that grounds a type-level h-motif contribution in a concrete instance. Equation (27) combines the type-level contribution from the prediction head with role weights from the convolutional branch and an instance match score. It selects a complete h-motif instance near the target node, and the three constituent hyperedges form the local explanation.

On the theoretical side, we use the definition of normalized h-motif frequencies to derive how hyperedge deletion changes the structural encoding. Proposition 3 bounds the changes in the frequencies of non-target types. Once the edited encodings are available, Theorem 1 uses finite differences of the quadratic head to calculate the exact logit increments from the structural and feature-prototype updates in the specified order. This traces how an edit to the original hypergraph affects the model’s output.

We revised the contribution statement in the introduction and the passages that describe and discuss the method. These revisions explain the distinct problems addressed by instance selection and intervention analysis and make the methodological and theoretical advances clearer.

AuthorsOpenReview ↗

Response to CW3

CW3:Scope of the N-SYNTH Results and Intervention-Effect Recovery

Response: Thank you for raising this concern. The original N-SYNTH experiment has a limitation as a test of intervention-effect recovery. The quadratic reference score was not used to generate the labels, so Table 2 measures agreement between the model’s response and changes in that reference score. To test whether IV-SENN can recover intervention effects defined by the label-generating mechanism, we added an experiment that uses the same nonquadratic rule to generate the training labels and compute the reference probability change for each test edit. IV-SENN was trained on unedited hypergraphs and evaluated on an independently generated test hypergraph. The results are shown below.

TargetEditsIV-SENN MAE ↓Baseline MAE ↓Dir. agreement ↑
hm8-target472 (460)0.0437 ± 0.00150.08620.8539 ± 0.0106
hm17-target465 (450)0.0392 ± 0.00160.06740.8018 ± 0.0183

Edits are reported as total (nonzero reference changes). The baseline predicts zero change.

For both edit targets, IV-SENN achieved lower probability-change MAE than the zero-change baseline and direction agreement above 0.80. These results show that, under this nonquadratic label-generating mechanism, IV-SENN can estimate the magnitude of probability changes caused by unseen hyperedge edits and recover their direction for most edits. Appendix D.11 (p. 29, lines 1544–1556) describes the protocol, and Table 3 in Section 5.4 reports the full results (p. 9, lines 446–452).

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

With the author response deadline approaching, we would appreciate any further feedback on whether our revisions have addressed your concerns.

We quantified how often candidate screening changes the final explanation, clarified the methodological and theoretical contributions, and added an experiment that uses the same nonquadratic rule to generate labels and define reference probability changes. The additional results support intervention-effect recovery under this known mechanism.

If these revisions and our responses have addressed your concerns, we would be grateful if you would consider a higher score. We remain available to clarify any outstanding questions.

Thank you again for your time and constructive comments.

Best regards, The Authors

Reviewer_4ambOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

IV-SENN makes node predictions on hypergraphs with an explicit quadratic model over 34 named concepts, and it uses no message passing in the prediction path. The concepts are 26 normalized h-motif frequencies sv\mathbf s_v, precomputed by MoCHy and fixed, and K=8K=8 learnable prototype matches fv\mathbf f_v, averaged over the mean/variance/pairwise statistics of the node's incident hyperedges. The class-cc logit is Fc(mv)=w~c⊤mv+mv⊤Qcmv+bcF_c(\mathbf m_v)=\tilde{\mathbf w}_c^\top\mathbf m_v+\mathbf m_v^\top Q_c\mathbf m_v+b_c with mv=[sv;fv]\mathbf m_v=[\mathbf s_v;\mathbf f_v]. Here w~c\tilde{\mathbf w}_c folds the linear and compositional weights into a single vector, and QcQ_c is symmetric. Because the head is explicit, the paper reads global explanations from the structural weights, computes local contributions ϕ\phi that sum exactly to the logit, and evaluates a structural edit by re-applying the fixed head to the recounted encoding (Thm. 1).

Critical strengths
  1. The logits are an explicit function of named, hypergraph-native concepts, and the local contributions sum exactly to the logit (Table 7). Placing h-motifs inside the decision path avoids the ambiguity that clique expansion introduces. This is the right design direction in my opinion.

  2. The paper states that the decomposition depends on the path order, that recounting lies outside the guarantee, and that the symmetric split of interaction terms is a convention. It also includes null-intervention controls. Fig. 3 reproduces from the stated counts, and Table 9 KL value recomputes exactly.

Critical weaknesses
  1. The structural explanations are not uniquely determined by the model's predictions. For every node with Tv>0T_v>0, the motif frequencies sv=mvstruct\mathbf s_v=\mathbf m^{\rm struct}_v sum to one, so many different weight settings give exactly the same logits. Adding a constant γ\gamma to all 26 structural weights of class cc and subtracting it from bcb_c changes nothing. Neither does adding u1⊤+1u⊤\mathbf u\mathbf 1^\top+\mathbf 1\mathbf u^\top to the structural interaction block while subtracting 2u2\mathbf u from the structural weights, since s⊤(u1⊤+1u⊤)s=2u⊤s\mathbf s^\top(\mathbf u\mathbf 1^\top+\mathbf 1\mathbf u^\top)\mathbf s=2\mathbf u^\top\mathbf s. This is an exact property of the encoding, not the incidental collinearity discussed after Prop. 6, and it undermines every explanation the paper reports. The intervention metric cannot detect any of this: a deletion changes sv\mathbf s_v by a vector that sums to zero, so the shift cancels. Table 2 shows exactly this. Removing ℓ1\ell_1 drops motif precision, while Pearson stays. The regularizer chooses the explanation, and the proposed verification is blind to that choice.

  2. Intervention verifiability holds trivially on real data and is circular on synthetic data. Thm. 1, F(x′′)−F(x)=[F(x′)−F(x)]+[F(x′′)−F(x′)]F(x'')-F(x)=[F(x')-F(x)]+[F(x'')-F(x')], holds for any function FF. Obtaining x′x' and x′′x'' requires recounting motifs and re-encoding features, both of which the paper places outside the guarantee. Theclosed-form estimate is an ordinary forward pass on the edited input, and on real data it equals the actual change by definition. The only non-trivial test, N-SYNTH, does not close the gap. The reference score SGT=3a−2b+8abS_{\rm GT}=3a-2b+8ab is written in the model's own input coordinates and has the head's linear-plus-pairwise form, so agreement shows only that the head can fit such a function. The labels are not generated from SGTS_{\rm GT}: the Bayes-optimal logit is a Poisson likelihood ratio, so absolute agreement in logit units (MAE 0.043) is unexplained. The pooled Pearson correlation mixes hm8- and hm17-deletion edits, whose effects have opposite signs, and this inflates the correlation.

Questions for authors
  1. The prediction path doesn't contain any message passing. It sees 34 node-level statistics. Classrooms are close to structurally exchangeable communities. What exactly are the input features (one for primary school, 16 for high school)? What accuracy do sv\mathbf s_v alone and fv\mathbf f_v alone achieve?

  2. What fraction of nodes on Cora/DBLP have ∣Ev∣=0|\mathcal E_v|=0 or Tv=0T_v=0? Such nodes receive the constant logit bcb_c. Are self-loop hyperedges added?

  3. Are the logit changes rescaled before computing MAE? Could you please report Pearson separately for hm8-target and hm17-target edits?

  4. Do Kc\mathcal K_c, the signs in Table 7, and the contribution shares survive a gauge-invariant parameterization (e.g., centered log-ratio)?

  5. How does Table 3 change if βv\beta_v is replaced by sv\mathbf s_v in Eq. 27? How do Fid$^-$ and GSACC compare against a random instance of the same selected type?

5 comments
AuthorsOpenReview ↗

Response to CW1

Thank you for taking the time to review our manuscript and for raising specific, constructive points. We hope the revised manuscript and our point-by-point responses below will assist you in reassessing the work.

CW1: Identifiability of Structural Attributions

Response: Thank you for your comments. We agree that normalizing the h-motif frequencies allows prediction-equivalent parameterizations to assign different contributions to individual motifs. For the same set of edits, identical logit changes yield the same intervention Pearson correlation even if the selected motifs differ. Proposition 6 establishes that symmetry removes the unobservable antisymmetric component of the interaction matrix. The transformations identified here exhibit an additional redundancy among symmetric parameterizations. Individual concept contributions are therefore defined for the fitted head under the stated interaction-allocation convention and can vary across prediction-equivalent parameterizations.

For a structural increment δ\boldsymbol{\delta} with the feature encoding held fixed, the quadratic head gives the exact logit change below.

ΔF=(ws+2Qsss+2Qsff)⊤δ+δ⊤Qssδ\displaystyle \Delta F = \bigl(\mathbf w_s + 2Q_{ss}\mathbf{s} + 2Q_{sf}\mathbf{f}\bigr)^{\top}\boldsymbol{\delta} + \boldsymbol{\delta}^{\top}Q_{ss}\boldsymbol{\delta}.

When both endpoint structural encodings sum to one, 1⊤δ=0\mathbf{1}^{\top}\boldsymbol{\delta} = 0, and the additional terms introduced by the two transformations cancel. If deletion destroys the last h-motif instance at a node, its structural encoding becomes zero. In this case, the first transformation leaves the pre-edit logit unchanged and shifts the post-edit logit by −γ-\gamma, so the logit increment also shifts by −γ-\gamma. The second transformation preserves the logits at both endpoints. More generally, equality of the logits at both endpoints of a specified edit implies equality of the predicted changes. In Table 2, Pearson measures agreement between predicted and reference-score changes, and motif precision measures the proportion of selected types that match the planted h-motifs.

Removing ℓ1\ell_1 regularization in Table 2 reduces motif precision from 0.967 to 0.082. Pearson correlation decreases from 0.948 to 0.935. The sparsity penalty can distinguish parameter representations with the same classification loss and thereby influence which motif types are selected. The ablation compares models trained with different objectives and shows that motif recovery is more sensitive to the penalty than the intervention Pearson correlation. The exact equivalence conditions in Appendix B.7 instead concern the logits obtained by applying specified parameter transformations to the same fitted head on the same inputs.

Appendix B.7, “Structural Parameter Equivalence and Intervention Invariance,” follows the proof of Proposition 6 (p. 16, lines 832–863). It gives both transformations and their effects on concept contributions and specifies when the logit change for the same edit is invariant, including edits involving a zero structural encoding.

AuthorsOpenReview ↗

Response to CW2 and Q1

CW2:Intervention Verification on Real and Synthetic Data

Response: Thank you for your comments. The sum of the two differences in Theorem 1 holds for any predictor. For our quadratic prediction head, Eq. (16) gives an exact expression for each finite increment once the edited encodings are available. Under the structure-first decomposition, the cross term between the two updates is assigned to the feature update. Motif recounting and feature re-encoding supply the edited inputs. On those inputs, the closed-form calculation agrees with the difference obtained from direct forward evaluation. The head timing in Table 7 starts once these encodings are available (Appendices B.1 and D.4).

In the original N-SYNTH experiment, class labels provide the prediction supervision, and the predefined quadratic score SGT=3a−2b+8abS_{\mathrm{GT}} = 3a - 2b + 8ab is used only for evaluation (Appendix C.1). As you point out, the reference score and the prediction head have compatible functional forms. The training objective does not use the reference scores or their changes as supervision, so compatibility alone does not ensure agreement on the test edits. Table 2 measures that agreement empirically.

The MAE of 0.043 in Table 2 directly compares the structural logit increment with the reference-score increment on their original scales (Appendix D.2). The label likelihood constrains the binary log-odds F1−F0F_1 - F_0. Adding the same input-dependent shift to both logits preserves the log-odds and can change a single-class logit increment. The reported MAE therefore measures how closely the trained head’s structural response follows this specified reference. In the additional experiment reported in Section 5.4 (Table 3, p. 9, lines 446–452), the same nonquadratic rule generates the labels and defines the reference probability change for each edit. This allows us to assess how well the model recovers probability changes under a known label-generating mechanism. Appendix D.11 describes the protocol.

To examine how pooling the two types of edits affects the interpretation of the correlation, we computed Pearson separately for each target motif. The pooled correlation was 0.948 ± 0.019, and the correlations for hm8-target and hm17-target edits were 0.921 ± 0.025 and 0.900 ± 0.029, respectively. Both groups retained strong positive correlations after grouping, with lower values than in the pooled analysis. The full breakdown appears in our response to Q3 and in Appendix D.4, Table 8.

Q1:Contact Dataset Features and Single-Encoding Accuracy

Response: Thank you for your comments. We verified that Contact-primary-school and Contact-high-school both use the 16-dimensional node features supplied with the public datasets. The dataset documentation does not specify what each feature dimension represents. The original Appendix C.2 incorrectly listed one input feature for Contact-primary-school. We corrected this to 16 in the revised Appendix C.2 (p. 18, lines 921–923).

As you point out, the role convolution does not enter the prediction path. The 34-dimensional node encoding still summarizes the target node’s neighborhood. The structural encoding sv\mathbf s_v contains normalized local h-motif counts, and the feature encoding fv\mathbf f_v aggregates prototype matches based on feature statistics computed across nodes in its incident hyperedges.

Following your suggestion, we evaluated models using only the structural encoding sv\mathbf s_v or only the feature encoding fv\mathbf f_v. The results are shown below. We used the same data splits and evaluation protocol as for the full model.

DatasetFull model ACCStructure-only ACCFeature-only ACCPaired ACC gain over feature-only
Contact-primary-school0.88140.7807 ± 0.05120.8310 ± 0.0557+0.0504 ± 0.0185
Contact-high-school0.93270.8598 ± 0.01590.8962 ± 0.0347+0.0366 ± 0.0135

The full model achieved the highest mean ACC on both datasets. In the paired comparison with the feature-only variant, adding structural encoding improved ACC by 0.0504 on Contact-primary-school and 0.0366 on Contact-high-school. Appendix D.10 (pp. 28–29, lines 1509–1532) describes the setup, and Table 14 (p. 29, lines 1535–1540) reports the results.

AuthorsOpenReview ↗

Response to Q2-Q4

Q2:Isolated Nodes, Zero Motif Counts, and Self-Loops

Response: Thank you for your comments. In the DHG coauthorship datasets, 320 of 2,708 Cora nodes (11.82%) and 0 of 41,302 DBLP nodes (0.00%) have no incident hyperedges (∣Ev∣=0\lvert\mathcal E_v\rvert = 0). The total h-motif count is zero (Tv=0T_v = 0) for 704 of 2,708 Cora nodes and 3,289 DBLP nodes (7.96%). These totals include nodes with no incident hyperedges.

We evaluated IV-SENN on the incidence structure supplied with each dataset and did not add singleton self-loop hyperedges. IV-SENN computes a node’s feature-prototype encoding from its incident hyperedges. Adding a singleton hyperedge for an isolated node would give the predictor access to that node’s own features through the new hyperedge and could affect the evaluation results. With the original incidence structure, both the structural encoding sv\mathbf s_v and the feature-prototype encoding fv\mathbf f_v are zero for an isolated node, so its logit for class cc equals the corresponding bias bcb_c.

When Tv=0T_v = 0 and ∣Ev∣>0\lvert\mathcal E_v\rvert > 0, the structural encoding is zero and the feature-prototype encoding is still computed from the incident hyperedges. The class logit then includes the bias, the linear feature term, and the feature–feature interaction term.

We added the counts and preprocessing details to Appendix C.2 (p. 18, lines 929–934) and clarified the two prediction cases in Appendix A.1 (p. 13, lines 692–695).

Q3:MAE Scaling and Pearson Correlation by Edit Target

Response: Thank you for your comments. We compute MAE using the original structural logit and reference-score increments without rescaling. To compute rMAE, we divide MAE by the mean absolute change in the reference score.

As requested, we report the Pearson correlations for hm8-target and hm17-target edits separately in the table below.

Target editEdits (five-seed total)Pearson (mean ± SD)
Pooled edits9310.948 ± 0.019
hm8-target4680.921 ± 0.025
hm17-target4630.900 ± 0.029

The table reports the mean and standard deviation across five seeds, with edit counts summed across seeds. Both edit types show strong positive correlations when evaluated separately. Appendix D.4 (p. 25, lines 1329–1332) describes the calculation, and Table 8 (p. 25, lines 1334–1340) reports the results. Section 5.4 (p. 8, lines 419–420) refers to the table.

Q4:Invariance of Selected Motif Types, Contribution Signs, and Shares

Response: Thank you for your comments. Regarding the invariance of the three explanation quantities, the second parameter transformation in our response to CW1 shows that all three can change. That transformation preserves the class logits for every admissible structural encoding, including the zero vector.

First, the transformation changes the effective structural linear weights used for thresholding and ranking in Eq. (37), so the selected set Kc\mathcal K_c can change. Second, for Cora node 785 in Table 9 (originally Table 7), the signed contribution of any one of the three reported structural motifs, hm8, hm17, and hm20, can be reversed by choosing a transformation along that motif’s coordinate. The node’s logit, feature-prototype contributions, and role evidence remain the same under this transformation. Third, the transformation transfers an equal and opposite amount between the structural linear and interaction terms. Their sum is preserved, while the reported shares, which are computed from the absolute magnitudes of the terms, can change.

This construction provides an exact test of invariance in the original motif-frequency coordinates, including when some or all frequencies are zero. For the same specified edit, the two equivalent heads agree at both endpoints and therefore give the same logit change. Appendix B.7 (p. 16, lines 850–863) now gives these consequences explicitly.

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

This paper proposes IV-SENN, a self-explaining hypergraph neural network that integrates h-motif structural concepts and feature prototypes through an additive concept-interaction head. It supports global and local explanations and verifies prediction changes under structural interventions. Experiments demonstrate promising classification performance and explanation quality.

Critical strengths

S1. The paper addresses an important problem by incorporating higher-order structural information into interpretable hypergraph learning. S2. The framework provides a unified approach to global, local, and intervention-based explanations. S3. The experiments demonstrate strong explanation fidelity and competitive predictive performance, supported by theoretical analysis and ablation studies.

Critical weaknesses

W1. Table 3 lacks an ablation removing the h-motif structural encoding or replacing it with simpler structural statistics, making its individual contribution difficult to assess. W2. The role-aware convolution is trained to approximate normalized h-motif frequencies (Eq. 10), but its advantage over directly using these frequencies for local explanation is not evaluated. W3. The instance-grounding strategy in Eq. 27 is not compared with simpler selection strategies, leaving its effectiveness in identifying the most influential local h-motif instances insufficiently validated. W4. Table 1 evaluates explanation methods with different prediction backbones, which limits the comparability of their explanation fidelity results.

Questions for authors

Q1. Could the authors include an ablation without h-motif structural encoding or replace it with basic structural statistics to isolate its contribution? Q2. How does the learned role distribution improve local explanations compared with directly using normalized h-motif frequencies? Q3. Could the authors compare the proposed instance-grounding strategy with random or contribution-based motif selection under the same explanation budget?

4 comments
AuthorsOpenReview ↗

Response to CW1–CW2

We thank the reviewer for the careful reading of our ablation and baseline experiments. We agree that additional controls were needed to assess the structural encoding, the local explanation rules, and the effect of the prediction backbone on fidelity. In response to your comments, we have refined the theoretical discussion and experimental analysis. We address each of your concerns below.

CW1:Contribution of the h-Motif Structural Encoding

Response: Thank you for pointing out the need for this control. On DBLP, we replaced the 26-dimensional h-motif structural encoding with five basic statistics: the number of incident hyperedges and the mean, standard deviation, minimum, and maximum of their sizes. The results are shown below.

Structural encodingGSACC ↑Fid⁺ ↑Fid⁻ ↓Expl.Ent.Eff.Mot.WCSRel.Drop ↓
h-motif encoding0.85310.82860.00930.32736.50.7398—
Basic structural statistics (5D)0.79000.77100.0700———7.40%

Expl.Ent., Eff.Mot., and WCS require h-motif contributions or weights. The dashes in these columns indicate that the metrics are inapplicable to the five-dimensional variant. This experiment confirms that with h-motif encoding, retaining the explanation subgraph preserves the model’s original prediction more often, and removing it changes the predicted class more often.

We added this comparison to the ablation table (p. 9, lines 453–458) and discuss it in Section 5.5 (p. 9, lines 474–475). Appendix C.2 (p. 18, lines 936–939) defines the five statistics.

CW2:Effect of Learned Role Weights on Local Explanations

Response: Thank you for asking what the learned role distribution adds to local explanations. The empirical distribution π records the relative frequencies of h-motif types around a node. The role-aware convolution estimates β from the structure and features of the node’s incident hyperedges, with π as its training target in Eq. (10). In Eq. (27), each candidate’s score is the product of its type’s role weight and the remaining score terms. If β equals π, every candidate receives the same score under either weighting. When the relative weights differ, they can change the ranking between candidates of different types and thus change which motif instance is selected. Within a type, candidates share the role weight and the class contribution factor. When their product is positive, the instance match L determines the order.

On DBLP, we compared learned β, direct use of π, and a control without role weighting. We held the predictor fixed during evaluation and retained the class contribution and instance match L in all three settings.

Role weighting on DBLPGSACC ↑Fid⁺ ↑Fid⁻ ↓Fid_KL ↑
Learned β (full rule)0.85310.82860.00931.0371
Empirical frequencies π0.83180.80790.03101.0112
No role weighting, with L retained0.81900.79550.04500.9956

Using empirical frequencies instead of omitting role weights raised GSACC by 0.0128. The learned β improved all four metrics over direct frequency weighting. It raised GSACC by a further 0.0213 and reduced Fid⁻ by 0.0217, so retaining only the selected explanation preserved the predictor’s original class more often.

Section 5.5 (p. 9, lines 476–479) summarizes the comparison. The first three rows of Table 12 in Appendix D.8 (p. 28, lines 1490–1495) report these results.

AuthorsOpenReview ↗

Response to CW3–CW4

CW3:Validation of the Instance-Grounding Rule

Response: Thank you for pointing out the need for simpler baselines to evaluate Eq. (27). On DBLP, we compared the full grounding rule with two alternatives using a fixed predictor and the same budget of one complete h-motif instance, consisting of three hyperedges, per eligible node. The first baseline samples uniformly from all candidate instances. The second selects the motif type with the largest positive class contribution, applies the same absolute-contribution fallback as the full rule when needed, and then samples uniformly within that type. The full rule scores candidates using the class contribution, learned role weight, and instance match.

Selection strategy on DBLPGSACC ↑Fid⁺ ↑Fid⁻ ↓Fid_KL ↑
Full grounding rule0.85310.82860.00931.0371
Uniform random selection from all candidates0.74530.73000.12850.7213
Contribution-based type selection, then uniform sampling within the type0.82940.78300.04170.9170

As shown in the table, the full rule achieved a GSACC gain of 0.1078 over uniform random selection and 0.0237 over contribution-based selection. Its selected instances preserved the predictor’s original class more often when retained and changed the predicted class more often when removed. This comparison shows that the full scoring rule improves instance selection on DBLP under the same explanation budget.

We summarize the results in Section 5.5 (p. 9, lines 476–479) and report the selection protocols and full results in Appendix D.8 (p. 28, lines 1483–1498).

CW4:Fidelity Comparisons Using a Shared Predictor

Response: Thank you for highlighting this issue. To address your concern that the prediction backbone can affect measured explanation fidelity, we added a comparison under a shared predictor. The graph self-explaining methods generate explanations within their prediction architectures, so we retain those architectures to preserve each method's explanation mechanism.

The table below reports results for three post-hoc hypergraph explainers (HyperEX, SHypX, and CF-Hyper) on four datasets, using the same HGCN for all three explainers on each dataset.

MethodDatasetGSACC ↑Fid⁺ ↑Fid⁻ ↓Fid_KL ↑
HyperEXCora0.71610.68430.07940.8744
DBLP0.79480.67350.09270.6235
Contact-high-school0.65470.55260.11540.8793
Contact-primary-school0.63950.56170.11210.8452
SHypXCora0.72220.70550.08260.9013
DBLP0.77060.63800.11740.6008
Contact-high-school0.68200.61300.09560.8236
Contact-primary-school0.67130.60400.09740.7730
CF-HyperCora0.7418———
DBLP0.7021———
Contact-high-school0.5721———
Contact-primary-school0.5512———

On each dataset, the three explainers share the same HGCN and therefore have identical ACC and F1. To avoid repeating these values in Table 1, we report the shared-HGCN results separately in Appendix D.9, Table 13 (p. 29, lines 1513–1527), with a reference in Section 5.2 (p. 7, lines 359–361).

AuthorsOpenReview ↗

Response to Q1-Q3

Q1:Replacing h-Motif Encoding with Basic Structural Statistics

Response: Thank you for suggesting this ablation. On DBLP, we replaced the 26-dimensional h-motif encoding with five basic structural statistics: the number of incident hyperedges and the mean, standard deviation, minimum, and maximum of their sizes. After the replacement, GSACC decreased from 0.8531 to 0.7900 (a relative drop of 0.0740), Fid⁺ decreased from 0.8286 to 0.7710, and Fid⁻ increased from 0.0093 to 0.0700. With h-motif encoding, retaining the explanation subgraph preserved the original prediction more often, and removing it changed the predicted class more often. See our response to CW1 for a detailed analysis of the results.

Q2:Learned Role Weights vs. Empirical Frequencies

Response: Thank you for asking what the learned role distribution adds to local explanations. On DBLP, replacing the empirical frequencies with learned β increased GSACC by 0.0213 and reduced Fid⁻ from 0.0310 to 0.0093. In this comparison, the selected subgraphs preserved the predictor’s decisions more often. See the first three rows of Table 12 (Appendix D.8, p. 28, lines 1490–1495) and the three-way comparison in CW2.

Q3:Instance Selection under the Same Explanation Budget

Response: Thank you for suggesting this comparison under the same explanation budget. On DBLP, we compared the full grounding rule with uniform random selection from all candidate instances and contribution-based type selection followed by uniform sampling within the selected type.

Compared with uniform random selection, the full rule increased GSACC from 0.7453 to 0.8531 (0.1078) and reduced Fid⁻ from 0.1285 to 0.0093 (0.1192). Compared with contribution-based selection, it increased GSACC from 0.8294 to 0.8531 (0.0237) and reduced Fid⁻ from 0.0417 to 0.0093 (0.0324). See our response to CW3 for the selection protocols and the full comparison.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

With the author response deadline approaching, we would appreciate any further questions or concerns you may have about our revisions.

In response to your feedback, we added a structural-encoding ablation, comparisons of role weighting and instance selection, and experiments with a shared HGCN predictor. These results provide additional evidence for the contribution of h-motif encoding and the effectiveness of the local explanation rules.

If these additions and our responses have addressed your concerns, we would be grateful if you would consider a higher score. We would be happy to clarify any remaining points while the discussion is open.

Thank you again for your time and constructive feedback.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.