{"id":"b5856d96-2afd-4f78-871a-b5dbfd72ddd6","arxiv_id":"2506.09866","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper introduces ELRUHNA, an unsupervised hypergraph alignment algorithm that jointly matches vertices and hyperedges, and reports accuracy gains of up to 25% over two baselines.","lead":"This paper introduces ELRUHNA, an algorithm that aligns two hypergraphs by jointly matching their vertices and their hyperedges using a bipartite representation. It reports accuracy improvements of up to 25% over existing methods on noisy real datasets and claims scalability to tens of thousands of vertices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol (self-noise, Poisson additions, 2-core filtering, undefined noise level) is not shown to represent real unsupervised hypergraph alignment; the 25% margin may not transfer.","rationale":"I agree with the reader's weakest assumption. The central claim is empirical, and its only evidence is the self-noise/2-core benchmark; the paper itself notes the 2-core is needed because degree-1 vertices are undifferentiable. That is a limitation for applications involving partial overlap and low-degree nodes. The missing HyperAlign baseline, absent code/data, and pseudocode ambiguities (e.g., X' used before assignment in Algorithm 1) are real but secondary: they affect verification, while the protocol concern affects whether the reported numbers would transfer to the problem the paper motivates. A structured-noise rerun would settle this without requiring new theory. Therefore the CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":10466,"tokens_out":10344,"duration_ms":128875,"concrete_test":"Release the exact noise-level schedule, then rerun the small-instance comparison with a structured noise model that deletes a fraction of original hyperedges and adds random hyperedges whose sizes follow the dataset's own size distribution, and evaluate both with and without 2-core filtering (or on held-out cross-network pairs). If ELRUHNA's 25% accuracy margin over BiG-Align/Cone-Align does not persist under this protocol at matched noise levels, the headline claim is protocol-dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim ('higher alignment accuracy ... up to 25%') is supported only by the Section 4 protocol: align a hypergraph to a noisy copy of itself, where noise consists of added random hyperedges with Poisson-distributed sizes (rate lambda = k-bar), and where every instance is first reduced to the 2-core of its bipartite representation. This is a favorable regime in three ways. (1) Query and data are the same hypergraph up to permutation plus additive noise; there are no deletions, partial overlaps, or scale differences, which are central to the cross-network applications named in the introduction. (2) The 2-core removes degree-1 vertices, i.e., exactly the vertices whose matches are hardest to determine from topology and on which a topology-only method cannot be evaluated. (3) The noise level is never defined: the text gives lambda for hyperedge size but no rule for how many random hyperedges are added per noise level, so the x-axes of Figs. 2 and 3 are not reproducible. If real noise is structured/correlated or includes degree-1 nodes, the observed >90% accuracy and 25% margin may be artifacts of this construction rather than a general improvement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ELRUHNA, an unsupervised hypergraph alignment algorithm that works on the bipartite incidence representation. The central formulation is the incidence alignment objective (Eq. 3), a binary quadratic program that simultaneously optimizes vertex and hyperedge matchings via the Frobenius inner product of normalized incidence matrices. The algorithm initializes similarities using left and right singular vectors of a block matrix derived from normalized incidence matrices and importance weights, then propagates them using a local matching rule and a cooling rule adapted from ELRUNA. Experiments compare ELRUHNA with BiG-Align and Cone-Align on real-world hypergraphs after taking the 2-core of the bipartite representation, using noisy copies of each hypergraph as test instances. The paper claims up to 25% accuracy improvement and scalability to tens of thousands of vertices in the bipartite representation.","tokens_in":10683,"tokens_out":5818,"duration_ms":62344,"significance":"If the central claims held, the paper would make a useful contribution: the incidence alignment objective gives a principled quadratic-program formulation for hypergraph alignment that naturally extends bipartite graph alignment, and the elimination-rule propagation with generalized eigenvector centrality initialization is a plausible topology-only heuristic. The use of a bipartite representation allows the method to scale beyond dense clique-expansion approaches. However, the evidence presented does not currently support the general claims in the abstract. The evaluation is confined to one favorable noise model, the closest unsupervised hypergraph baseline is absent, the algorithm pseudo-code has undefined variables, and the code/data are not available for verification. These are correctable but load-bearing gaps, so the significance is conditional on a substantially revised evaluation and a fully specified algorithm.","major_comments":[{"comment":"The evaluation used for the headline 'up to 25%' claim is only self-alignment under additive Poisson-sized random hyperedges, after reduction to the 2-core of the bipartite representation. The number of added hyperedges per noise level is never defined, so the x-axes of Figs. 2 and 3 are not reproducible. Furthermore, this protocol excludes deletions, partial overlaps, and degree-1 vertices, and therefore does not test the cross-network scenarios (social networks, bioinformatics) described in the introduction. Please define noise level precisely, add deletion and partial-overlap experiments, and report results without the 2-core filter or at least quantify the sensitivity to that filter.","section":"Section 4, Figs. 2 and 3"},{"comment":"HyperAlign [9] is the most closely related unsupervised hypergraph alignment method, and the paper adopts its datasets and 2-core protocol, yet it is omitted from all comparisons because of 'reproducibility issues.' Without this baseline, the claim of state-of-the-art quality in the abstract is not supported. Please either provide a comparison to HyperAlign (e.g., via the authors' released code or an independent implementation) or restrict the claim to the compared baselines.","section":"Section 4, first paragraph"},{"comment":"In Algorithm 1, the assignment 'Y ← W'E + B'Q X' (B'D)^T' uses X' before it is defined; later 'Y' ← dominant match(Y')' and 'X' ← dominant match(WV + (B'Q)^T Y' B'D)' use X' and Y' without initialization or definition. Also, Algorithm 2 is called without specifying which similarity matrices it updates. As written, the algorithm cannot be reimplemented. Please rewrite the pseudo-code with explicit initialization of X' and Y' and clear input/output for Algorithm 2.","section":"Algorithm 1"},{"comment":"The generalized eigenvector centrality initialization in Eq. (4) contains undefined notation: the vector u_Q is not defined, and the block matrix and the bracketed vectors appear to have inconsistent dimensions (the top-left block W_V has size |V_Q|×|V_D|, while the bottom-right block W_E^T has size |E_D|×|E_Q|). Please define all variables and give the exact dimensions of each block so that the objective is well-posed.","section":"Section 3.2, Eq. (4)"},{"comment":"The reproducibility statement says 'Code and data are available at [link will be added upon acceptance],' but no code or data are provided with the submission. Given the number of free parameters (niter, beta, sparsification k, t1, t2, noise level) and the current pseudo-code gaps, the experiments cannot be independently verified. Please provide code/data or a detailed supplement for review.","section":"Reproducibility statement"}],"minor_comments":[{"comment":"The domain of the optimization in Eq. (1) is written 'σ: V^D ,→ V^Q', which appears to be a typo for σ: V^Q → V^D; please correct.","section":"Section 2.1"},{"comment":"'dominant match' is used before being formally defined; the reference to [11] is helpful, but a precise definition would make the rule self-contained.","section":"Section 3.3, Eq. (6)"},{"comment":"The constraint notation 'X⊙Π_V = X' and 'Y⊙Π_E = Y' is not explained; Π_V and Π_E are never defined.","section":"Section 3.1, Eq. (3)"},{"comment":"The phrase 'non-convergent but stabilized Gauss-Seidel relaxation' is unclear and seems contradictory; please clarify or remove it.","section":"Section 3.2"},{"comment":"The caption and text do not specify whether |V| and |E| refer to the original hypergraph or the 2-core-reduced instance; please state this explicitly.","section":"Table 1"},{"comment":"The text reports standard deviation up to 0.3 at higher noise levels but does not report error bars or per-instance standard deviations in Fig. 2; adding confidence intervals would strengthen the comparison.","section":"Section 4.1, Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's scope and method are within the journal's interests, but the evaluation and algorithm specification need substantial revision before the claims can be assessed. In particular, the missing HyperAlign comparison and the incompletely specified noise injection protocol are concerns that should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the incidence alignment objective in Eq. 3. Jointly optimizing vertex and hyperedge matchings with edge-size-normalized weights is a natural improvement over hyperedge correctness, and the paper honestly notes it is a weighted extension of BiG-Align. The ELRUHNA heuristic then propagates similarities with local matching and cooling, which is a sensible adaptation of the ELRUNA graph-alignment work.\n\nThe experiments show the method working well on noisy copies of real hypergraphs, with up to 25% over BiG-Align and >90% accuracy in low noise. That is a real result on the tested instances, and the paper admits the clique-expansion baseline struggles. It also scales to tens of thousands of bipartite vertices, which is useful.\n\nThe soft spots are in the evaluation protocol. Every instance aligns a hypergraph to a noisy version of itself, where noise is purely additive random hyperedges. There are no deletions, partial overlaps, or scale differences, so the 25% margin has not been shown to transfer to the cross-network applications named in the introduction. The 2-core preprocessing removes degree-1 vertices, exactly the ones a topology-only method should struggle with. The noise level is also under-specified: the paper defines lambda for hyperedge size but not how many random hyperedges are added, so the x-axis in Figs. 2 and 3 is not reproducible. Those are fixable in revision, but they do limit what the headline claim supports right now.\n\nHyperAlign, the closest unsupervised hypergraph baseline, is omitted with a reproducibility note. That is honest, but it means the comparison is against graph methods adapted to hypergraphs, not recent hypergraph-specific work. The missing code and data are a bigger problem: \"link will be added\" is not enough for the 25% claim to be independently checked.\n\nThe math is coherent and the paper is well written. The pseudocode has a few ambiguities (for example, the role of X* inside the while loop), but nothing fatal. The authors are also honest about the high-noise instability.\n\nThis paper is for researchers working on unsupervised hypergraph alignment, especially those who care about topology-only matching. It deserves a serious referee: the formulation is new and the empirical direction is sensible. I would send it to review, but with a request that the authors provide code and data, either add HyperAlign or explain why it cannot be run, and broaden the evaluation beyond self-noise with, say, a deletion or partial-overlap test.","headline":"New incidence-based objective and heuristic for unsupervised hypergraph alignment, but the evaluation is built on a favorable self-noise protocol and the code is missing.","tokens_in":11238,"tokens_out":2477,"would_cite":false,"duration_ms":28877,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C65","90C20","68R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper presents ELRUHNA, a rule-based solver that aligns hypergraphs by matching vertices and hyperedges together, reporting up to 25% higher accuracy than prior methods on real-world instances.","keywords":["hypergraph alignment","incidence alignment","bipartite representation","elimination rules","similarity propagation","unsupervised alignment","network alignment"],"falsifier":"To test the claim, run ELRUHNA and BiG-Align on a real paired dataset with known vertex correspondences, keeping degree-1 vertices and adding correlated, not independent, noise; if the accuracy gap shrinks to zero or reverses, the reported 25% gain depends on the 2-core and Poisson-noise assumptions.","tokens_in":10266,"feed_emoji":"🔗","tokens_out":9337,"duration_ms":86020,"temperature":0.7,"pith_summary":"This paper tries to establish that unsupervised hypergraph alignment, an NP-hard problem, can be solved accurately and at scale by working on the bipartite representation and aligning vertices and hyperedges together rather than vertices alone. It introduces the incidence alignment objective, a binary quadratic optimization problem, and an iterative solver, ELRUHNA, that propagates similarities through elimination rules. The paper reports that ELRUHNA outperforms state-of-the-art methods by up to 25% in alignment accuracy on real-world hypergraphs, and handles instances with tens of thousands of vertices in the bipartite representation. If true, this would give an unsupervised, topology-only tool for entity correspondence in high-order relational data such as co-authorship, contact, and email networks.","feed_headline":"Aligns noisy hypergraphs 25% more accurately by matching edges too","feed_subtitle":"A rule-based solver recovers vertex and hyperedge correspondences together, beating BiG-Align on real hypergraphs.","key_machinery":"The mechanism is the incidence alignment formulation on the bipartite representation of hypergraphs, where the incidence matrix $B_{ue} = 1[u \\in e]$ is normalized by hyperedge size to $\\tilde B_{ue} = B_{ue}/\\sqrt{|e|}$. The objective couples vertex and hyperedge match matrices $X, Y$ through the term $\\langle \\tilde B_Q Y, X \\tilde B_D\\rangle_F$, so a high score requires that matched vertices sit in matched hyperedges. ELRUHNA propagates similarities with two elimination rules: a local matching rule that aligns the neighborhoods of a candidate pair and takes the best dominant match, and a cooling rule that decays non-dominant similarities toward zero. Initial similarities come from generalized eigenvector centrality, computed as the leading singular vectors of a block matrix that incorporates known match weights $W_V, W_E$.","core_discovery":"The central claim is that aligning a hypergraph to a noisy copy of itself can be done more accurately by solving incidence alignment: simultaneously finding an injection of query vertices and an injection of query hyperedges so that incident pairs map to incident pairs. The paper shows that this objective, written as $\\langle \\tilde B_Q Y, X \\tilde B_D\\rangle_F$ with normalized incidence matrices, generalizes the BiG-Align bipartite objective and is a quadratic program for rank-2. ELRUHNA solves it by iterating two elimination rules, local matching and cooling, over the similarity matrices, and by initializing similarities from left and right singular vectors of the normalized incidence matrix. Reported experiments on small and large real hypergraphs show over 90% average true accuracy in low noise for some instances, up to a 25% improvement over BiG-Align, with weak clique-expansion-based baselines.","pith_inferences":["A testable extension is to run ELRUHNA on paired real hypergraphs with ground-truth correspondence and structured noise, such as missing or merged hyperedges; if the reported gain shrinks, the independent-Poisson noise model is the source of the advantage.","The incidence alignment objective points toward a hyperedge-first variant: when vertex identities are ambiguous, aligning hyperedge neighborhoods first could seed the vertex matching, reversing the paper's vertex-first initialization.","Because the solver matches vertices and hyperedges in one objective, it may also serve as a building block for temporal or multilayer hypergraph alignment, where the same entities appear across snapshots with evolving hyperedge memberships.","Adapting the cooling threshold to the noise level is a concrete follow-up suggested by the reported instability at high noise, and could improve robustness without changing the objective."],"forward_implications":["If the claim holds, unsupervised alignment of real high-order networks no longer requires node attributes or embeddings: topology alone, via the bipartite representation, is enough to recover most correspondences under mild noise.","The incidence alignment objective unifies hypergraph alignment with bipartite graph alignment, so algorithmic advances for one transfer to the other.","The sparse mode, keeping only the top $\\lceil \\log_2(\\max(|V|,|E|))\\rceil$ candidate pairs, lets the method handle tens of thousands of vertices, making it a practical heuristic for large pattern-discovery tasks.","The accuracy gap over clique-expansion baselines indicates that preserving higher-order incidence, rather than pairwise proximity, is what drives alignment quality on noisy hypergraphs."],"supporting_citations":[{"why":"It supplies the elimination-rule propagation scheme and the comparison function that ELRUHNA adapts.","marker":"[20]"},{"why":"It provides the bipartite graph alignment objective that incidence alignment generalizes and acts as the main baseline.","marker":"[12]"},{"why":"It provides the embedding-based baseline and the KD-tree sparsification strategy used for large instances.","marker":"[8]"},{"why":"It motivates the bipartite-representation approach and supplies the evaluation protocol, including the 2-core and noise settings.","marker":"[9]"},{"why":"It supplies the dominant-match approximate matching routine used inside the local matching update.","marker":"[11]"},{"why":"It provides the random hypergraph noise model used to create the noisy query copies in the experiments.","marker":"[2]"},{"why":"It supplies the real-world hypergraph datasets used in the evaluation.","marker":"[13]"}],"fun_headline_variants":["Joint vertex-edge matching boosts hypergraph alignment accuracy","Align hypergraphs by matching vertices and hyperedges together","ELRUHNA: rule-based hypergraph alignment beats BiG-Align on real data","Incidence alignment improves hypergraph correspondence by up to 25%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that aligning a hypergraph to a noisy copy of itself, with independent random hyperedges and only vertices and hyperedges that appear in at least two interactions kept, is representative of real unsupervised hypergraph alignment tasks.","fun_headline_variants_meta":{"raw":{"variants":["Joint vertex-edge matching boosts hypergraph alignment accuracy","Align hypergraphs by matching vertices and hyperedges together","ELRUHNA: rule-based hypergraph alignment beats BiG-Align on real data","Incidence alignment improves hypergraph correspondence by up to 25%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1409,"prompt_tokens":884,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":452}},"tokens_in":500,"tokens_out":525,"duration_ms":6250,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:38:31.079671+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"To test the claim, run ELRUHNA and BiG-Align on a real paired dataset with known vertex correspondences, keeping degree-1 vertices and adding correlated, not independent, noise; if the accuracy gap shrinks to zero or reverses, the reported 25% gain depends on the 2-core and Poisson-noise assumptions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the elimination-rule propagation scheme and the comparison function that ELRUHNA adapts."},{"cited_title":"In: 2013 IEEE 13th international conference on data mining, IEEE (2013) 389–398","cited_arxiv_id":null,"evidence_quote":"It provides the bipartite graph alignment objective that incidence alignment generalizes and acts as the main baseline."},{"cited_title":"In: Proceedings of the 29th ACM International Conference on Information & Knowledge Manage- ment","cited_arxiv_id":null,"evidence_quote":"It provides the embedding-based baseline and the KD-tree sparsification strategy used for large instances."},{"cited_title":"In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining","cited_arxiv_id":null,"evidence_quote":"It motivates the bipartite-representation approach and supplies the evaluation protocol, including the 2-core and noise settings."},{"cited_title":"In: SC ’12: Proceedings of the International Conference on High Performance Computing, Networking, Stor- age and Analysis","cited_arxiv_id":null,"evidence_quote":"It supplies the dominant-match approximate matching routine used inside the local matching update."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the random hypergraph noise model used to create the noisy query copies in the experiments."},{"cited_title":"Journal of Open Source Software8(85) (2023) 5162","cited_arxiv_id":null,"evidence_quote":"It supplies the real-world hypergraph datasets used in the evaluation."}],"review_version":1}