acceptodds
Withdrawn from ICLR 2027

HyperCF: Bidirectional Incidence Counterfactuals for Hypergraph Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hypergraphs model higher-order interactions that pairwise graphs cannot represent, and hypergraph neural networks (HGNNs) are increasingly used to learn over them. As these models enter decision settings, structural counterfactual explanations become important: they ask for a minimal incidence edit under which the prediction would change. Existing hypergraph counterfactual methods are limited to deletion-only interventions on the original incidence structure, so the search fails on many targets. We expand the intervention space to bidirectional incidence editing, allowing a target both to leave incidences and to join existing hyperedges, and present HyperCF , which controls this expansion with a hop-bounded additive candidate space, learning-rate calibration based on deletion gradients with a coupled additive rate, and symmetric redundancy pruning. Out of seven widely adopted benchmarks, HyperCF raises counterfactual success over deletion-only search on six datasets at a two-hop radius and five at a one-hop radius, with gains of up to 40 percentage points, while the mean explanation size remains comparable and often smaller.

Primary area:
interpretability and explainable AI
Keywords:
hypergraph neural networks, counterfactual explanation, explainability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ib9wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes structural counterfactual explanations for node classification with hypergraph neural networks. It argues that the only existing hypergraph counterfactual method, CF-HyperGNNExplainer, is limited to deleting incidences, so it can only sparsify a target node’s original memberships. The authors propose HyperCF, which lets a target node both leave its current hyperedges and join existing nearby hyperedges through a single soft incidence mask. To make this search tractable, HyperCF restricts additions to nearby hyperedges, uses a smaller learning rate for additions, and prunes redundant edits. On seven benchmarks, it improves success over the deletion-only baseline on most datasets with similar or smaller explanation sizes, but underperforms on DBLP-CA.

Critical strengths

1.Clear and well-motivated problem. The paper identifies a concrete limitation of existing hypergraph counterfactual methods: deletion-only search can only sparsify a node’s original memberships.

2.Clean ablation of the pruning step. The ablation toggles only the pruning step on the same node set and reports deletions and additions separately, which cleanly isolates its effect.

3.Native incidence-level design. HyperCF edits the incidence matrix directly without star expansion.

Critical weaknesses

1.Limited novelty. The method mainly transfers existing bidirectional editing and pruning ideas from pairwise graphs to hypergraph incidences, building on CF-HyperGNNExplainer for hypergraph counterfactual search, and the paper itself acknowledges in Section 1 that these ideas already exist for pairwise edges.

2.Lack of theoretical analysis. The paper does not analyze when joining existing hyperedges can flip a prediction that deletions cannot, so the benefit of additions is shown only empirically, and the failure on DBLP-CA is explained only by speculation.

3.Large gap between pruned and unpruned explanations. Without pruning, explanation sizes on Pubmed and Mushroom reach 23.26 and 29.93 edits, compared with 7.06 and 3.81 after pruning, and the authors should explain why the search itself produces so many redundant edits.

4.Plausibility of added incidences. The paper does not evaluate whether adding a node to an existing hyperedge yields a realistic or meaningful counterfactual, so it is unclear whether the gains reflect useful explanations or implausible edits.

5.Limited experimental reliability. Experiments use only one HGNN backbone and report no standard deviations, so it is unclear whether the method generalizes and whether small gains are significant.

Questions for authors

(1) Why are unpruned explanations so large on Pubmed and Mushroom?

(2) How do you ensure that added incidences are plausible?

(3) Can you report standard deviations and results on another HGNN backbone?

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

The paper addresses counterfactual explanations for node classification using hypergraph neural networks by identifying structural edits to the hypergraph that change a target node's prediction. Prior work on hypergraphs finds counterfactuals using only edge deletions. The paper proposes HyperCF, which searches over both deleting and adding the target node's incidences. It restricts additions to nearby hyperedges, balances the learning rates of the two edit types, and prunes redundant edits after the search.

Critical strengths

S1. Learning rates. Identifying the need for separate learning rates for deletions and additions is a good observation.

S2. Writing. The paper is well written and easy to follow.

Critical weaknesses

W1. Limited novelty. This paper is an application of existing counterfactual-explanation machinery to hypergraphs rather than a novel contribution.

  • The paper acknowledges that "bidirectional edits and post-hoc pruning already exist for pairwise edges" (C2Explainer, ATEX-CF), and each component of HyperCF has a close counterpart there.
    1. The joint mask that initialises existing incidences near 1 and candidate additions near 0 is part of C2Explainer's "supergraph as input" trick.
    2. The hop-bounded additive candidate set mirrors C2Explainer's receptive-field candidate edges and ATEX-CF's neighbourhood-restricted, budgeted candidates. ATEX-CF further ranks additions with an attack, which is more refined than a fixed hop radius.
    3. The loss, thresholding, and best-valid-counterfactual selection are inherited from CF-GNNExplainer/CF-HyperGNNExplainer.
    4. Greedy redundancy pruning is ATEX-CF's minimality step, which additionally ranks edits by gradient importance. The only new component is the deletion-only learning-rate calibration with the coupled ratio ρ\rho, a tuned heuristic.
  • The paper argues in the intro that going from standard graphs to hypergraphs, incidences change "the candidate geometry, the optimization balance, and the meaning of redundancy", hence methods are not transferrable from standard graphs to hypergraphs. I disagree.
    • An incidence (v,e)(v, e) is an edge in the star expansion, and the HGCN operator is normalised message passing on that bipartite graph, so HyperCF essentially applies a C2Explainer-style mask to star-expansion edges incident to the target.
    • Reverting a deletion or an addition on hypergraphs means the same as on graphs, so Algorithm 2 doesn't use anything hypergraph-specific. The imbalance between many addition candidates and few deletion candidates arises in any graph method that adds edges. The hypergraph-specific choices (joining only existing hyperedges, editing only the target's row) restrict the search space rather than add a technique. Conversely, HyperCF omits components those methods offer, e.g., ATEX-CF's plausibility loss and C2Explainer's straight-through binarisation.

W2. Incomparable baseline numbers. Section 5 states that the baseline HyperEX, SHypX, and CF-GNNExplainer results "are taken from" Veglianti et al. (2026), i.e., copied from the baseline tables (Tables 5 and 6) of the CF-HyperGNNExplainer paper, while CF-HyperGNNExplainer itself is rerun by the authors.

  • First, the copied numbers come from a different setup. According to Section 5, this paper trains new models for explanation, using new seeds and different node samples than those of Veglianti et al. Restating numbers from a published paper is only acceptable when the setups are identical.
  • Second, mixing copied and rerun rows within a single table is unacceptable.
  • The reproduced CF-HyperGNNExplainer is much weaker than what Veglianti et al. report for its NHP (node–hyperedge perturbation) variant, which is the variant reproduced here. Success rates, reproduced (Table 1) vs. originally reported (Veglianti et al., Table 5): Mushroom 0.60 vs. 0.96; DBLP-CA 0.76 vs. 0.88; Citeseer 0.91 vs. 1.00; Cora 0.89 vs. 0.94. Mushroom is also HyperCF's largest gain over the reproduced baseline (0.60 to 1.00 at h=2h=2), and on DBLP-CA the originally reported NHP (0.88) beats HyperCF at both radii (0.75 and 0.60).
  • Finally, Veglianti et al. evaluate on ten AllSet datasets; this paper drops Zoo, ModelNet40, and House. House (committee hyperedges) is one of the few naturally higher-order benchmarks and the hardest case for deletion-only search based on Veglianti et al.'s results, making it a key test.

As it stands, the empirical claims are not supported: the baselines need to be rerun under one protocol, on the full benchmark suite, before any comparison can be trusted.

W3. Missing additive baselines. All four baselines in Table 1 are deletion-only. Hence, the central claim, that adding incidences helps, isn't tested against another method that supports addition. Table 1 excludes the SOTA counterfactual techniques cited in the related work section (ATEX-CF and C2Explainer) on the grounds that they are designed for standard graphs rather than hypergraphs. However, CF-GNNExplainer, an older technique from 2022, is included as a baseline via the star expansion, even though it is designed for standard graphs. Therefore, the authors should similarly extend the more recent techniques, compare them, or otherwise provide justification.

W4. Counterfactual vs. Out-of-distribution. A counterfactual is a minimally perturbed input with a different predicted class label. However, the counterfactual must remain in-distribution. How does the proposed method ensure that the counterfactual doesn't go out-of-distribution? Posed another way: how do you differentiate your counterfactual edit from an adversarial attack?

W5. Direct connections only. The proposed method only considers the target node's direct connections to identify its counterfactual. This is suboptimal, as perturbing the neighbourhood can yield better counterfactuals. Finding counterfactuals in the combined space (direct connections + neighborhood) is a significantly harder task than the one tackled in the paper.

W6. Incomplete definition. Section 3.2 defines a structural counterfactual as "an edited incidence matrix H′\mathbf{H}' satisfying f(H′≠y^t)f(\mathbf{H'} \neq \hat{y}_t)". This definition doesn't enforce minimality.

W7. Figure 1. The coloured blobs and their overlap are hard to make out. Consider adding borders and/or hashing.

Questions for authors

Section 4.3.

  1. Algorithm 2 processes deletions before additions ("Δ−\Delta^- then Δ+\Delta^+"), but the order inside each group is not specified. Are edits taken in order of their mask values?
  2. What's wrong with applying greedy to Δ−∪Δ+\Delta^- \cup \Delta^+ ranked by mask scores?
Reviewer_wbRKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies counterfactual explanations for node classification with HGNNs. Unlike prior work, which can only delete the target node, HyperCF proposes to allow the target node to also join existing hyperedges.

Critical strengths
  • The question is well-posed and incidence addition is an interesting counterfactual for HGNNs
  • Additions seem impactful. Per Table 6, after pruning, most of the edits are additions.
Critical weaknesses
  • Baselines - CF-HyperGNNExplainer had much better success rates originally published than in Table 1. Baseline hyperparameters were not tuned. This is a significant difference, why?
  • Results consistency - Table 1 and Table 4, for example, contradict each other. No variance.
  • Only evaluation on one HGCN.
Questions for authors
  • Are added edges really plausible? For example, Figure 1 shows that two authors were replaced by a third, is this a plausible explanation?
  • How does HyperCF differ from a simple greedy search in performance?
  • What is the runtime per explanation?

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

Related papers

Discussion (0)

Sign in to comment.