acceptodds
Withdrawn from ICLR 2027

Dual-View Fuzzy Membership and Candidate-Adaptive Multi-Source Evidence Fusion for Hyperedge Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hyperedge prediction requires identifying missing higher-order relations from incomplete hypergraphs. Existing approaches typically encode candidate hyperedges into unified representations, implicitly combining heterogeneous formation cues and overlooking how different candidates rely on different types of evidence. However, candidate hyperedges may emerge from diverse formation patterns, driven by structural participation, attribute semantics, or their cross-view consistency. Therefore, prediction requires explicitly modeling these complementary formation evidences and adaptively determining their contributions for each candidate. To address this issue, we propose FAME-HP, a framework that models candidate hyperedges through dual-view fuzzy overlapping memberships and candidate-adaptive evidence fusion. FAME-HP learns structural and semantic membership spaces, derives candidate-level structural, semantic, and cross-view consistency evidence, and dynamically adjusts their contributions according to candidate-specific contexts. Experiments on six real-world hypergraphs under four evaluation protocols demonstrate that FAME-HP achieves the best or second-best performance across all six datasets in average AUROC and average precision, outperforming strong hyperedge prediction baselines on most benchmarks. These results demonstrate that explicitly organizing and adapting heterogeneous formation evidence provides an effective approach for identifying missing higher-order relations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hyperedge Prediction, Hypergraph Representation Learning, Fuzzy Overlapping Communities, Multi-View Learning, Adaptive Evidence Fusion, Hard-Negative Discrimination

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_gUenOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FAME-HP to address hyperedge prediction by separating three kinds of evidence for a candidate hyperedge: structural compatibility, attribute compatibility, and cross-view consistency. It represents nodes using fuzzy memberships in structural and attribute spaces, allowing a node to belong to multiple latent groups. A candidate-adaptive gating mechanism then learns how strongly each evidence source should influence the final prediction.

The method is evaluated on six real-world hypergraphs using four negative-sampling protocols and compared with several hyperedge-prediction baselines. FAME-HP achieves the best or second-best average AUROC and AP on all six datasets, with particularly strong gains on difficult one-node-replacement negatives. Ablation and sensitivity studies further show that fuzzy memberships, membership entropy, attribute information, and cross-view consistency each contribute to the performance.

Critical strengths
  1. The paper explicitly separates structural compatibility, attribute compatibility, and cross-view consistency, while using fuzzy memberships to retain overlapping latent affiliations. This is a meaningful distinction from much prior hyperedge-prediction work that primarily improves candidate aggregation, hypergraph encoding, or negative sampling, and it provides a reasonably interpretable way to understand why different candidate hyperedges may require different evidence.

  2. FAME-HP achieves first- or second-best average performance across the six evaluated datasets and is especially strong under CNS, where negatives differ from true hyperedges by only one node. The controlled ablations consistently support the value of fuzzy membership, membership uncertainty, attribute information, and cross-view consistency, making the empirical contribution consequential even though the individual modeling ingredients are not entirely new.

Critical weaknesses
  1. Candidate-specific higher-order interactions have already been modeled using self-attention in Hyper-SAGNN, and the role of different orders of structural information in hyperedge prediction has already been studied systematically [1, 2]. Overlapping node memberships have also been used directly for missing-hyperedge inference, while flexible multiset-based hypergraph representation learning has been established by AllSet [3, 4]. The proposed method may therefore be viewed primarily as a combination of established ideas: fuzzy memberships, structural and attribute views, cross-view agreement, and adaptive gating, rather than as a fundamentally new modelling principle.
  2. Although ablations are provided for fuzzy membership, membership entropy, the attribute view, and cross-view consistency, the adaptive gate is not compared against simple alternatives such as equal weighting or globally learned fixed weights. This is consequential because candidate-specific weighting is presented as a central contribution, yet its incremental benefit is not isolated from the benefit of adding the underlying evidence features. Moreover, sensitivity to the construction of negative hyperedges has previously been demonstrated for hyperedge prediction [5]. More fundamentally, sampled positive-versus-negative classification has been shown not necessarily to reflect the actual hyperedge-retrieval problem, with classification and prediction performance reported to be poorly correlated [6]. Stronger evidence is therefore needed to establish that the adaptive fusion mechanism itself drives the gains and that these gains extend beyond the selected sampled-negative protocols [5, 6].

References

[1] Hyper-SAGNN: A Self-Attention Based Graph Neural Network for Hypergraphs, ICLR'20

[2] How Much and When Do We Need Higher-order Information in Hypergraphs? A Case Study on Hyperedge Prediction, WWW'20

[3] Inference of Hyperedges and Overlapping Communities in Hypergraphs, NatCommun'22

[4] You Are AllSet: A Multiset Function Framework for Hypergraph Neural Networks, ICLR'22

[5] AHP: Learning to Negative Sample for Hyperedge Prediction, SIGIR'22

[6] Prediction Is NOT Classification: On Formulation and Evaluation of Hyperedge Prediction, ICDMW'24

Questions for authors
  1. How much of the reported improvement is specifically due to candidate-adaptive fusion? Could there be a controlled comparison against equal weighting and globally learned fixed weights, while keeping the same three evidence sources, base predictor, and training protocol?
  2. How were the fixed coefficients used to construct the structural, attribute, and cross-view evidence scores selected? Were these coefficients fixed before evaluating the six benchmark datasets, and was any validation or benchmark performance used in choosing them?
  3. How closely matched are the evaluation conditions for FAME-HP and the externally reported baselines? In particular, are the data preprocessing, train/validation/test splits, candidate sets, node features, and negative samples identical, or are only the nominal benchmark protocols matched?
Reviewer_kqorOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FAME-HP, which integrates the fuzzy membership relationship and candidate conditions through the fusion gating of dual views to construct the structure, attribute semantics, and cross-view consistency evidence for hyperedge candidates, and uses it as a correction term for the frozen base predictor. The method has been validated on six datasets. Although the overall design has certain rationality, the paper still has issues such as insufficient comprehensive literature review, less coherent writing logic, and lack of novelty in innovation.

Critical strengths

S1:The research topic is quite popular.

Critical weaknesses

W1: The problems existing in the current related methods are not explained clearly enough. It is suggested to provide typical examples. Additionally, the 2022 published paper (Contisciani et al., 2022) is too old and insufficient to support the corresponding viewpoints. W2: The connection between the research works of this paper and the existing related methods' problems is also not clear enough. It is recommended to reorganize the language to explain it clearly. W3: The idea is not particularly novel. It involves representing learning from three perspectives: structure, semantics, and considering both structure and semantics simultaneously. During the integration process, adaptive adjustments were made to the weights. This approach is quite common in graph learning-related papers and is not particularly novel. W4: Why is Overlapping communities and fuzzy memberships introduced in the related work? Does the proposed method in this paper directly adopt the existing hypergraph community detection method for node representation? W5: In the experimental part, the comparison methods for 2026-2027 are missing, resulting in insufficient persuasiveness of the experimental results. W6: The shortcomings of the proposed algorithm in this paper have not been analyzed. W7: The reference literature review is not comprehensive. The reference from 2026 is missing, which makes the analysis of the existing related methods' problems in the research motivation questionable. For example, the following paper is closely related to the research work in this paper. It is recommended to introduce it in the related work section and conduct a comparative analysis with the research work in this paper, as well as necessary experimental comparisons. [1]Ming Li, Zhanle Zhu, Xinyi Li, Lu Bai, Lixin Cui, Feilong Cao, Ke Lv: HyperNoRA: Hyperedge Prediction via Node-Level Relation-Aware Self-Supervised Hypergraph Learning. AAAI 2026: 23047-23054 [2]Ming Li, Huiting Wang, Yuting Chen, Lu Bai, Lixin Cui, Feilong Cao, Ke Lv: Self-Supervised Hypergraph Learning with Substructure Awareness for Hyperedge Prediction. AAAI 2026: 23055-23062 [3]Changyuan Tian, Li Jin, Zequn Zhang, Zhicong Lu, Wen Shi, Jianhua Yin, Shiyao Yan, Zhi Guo: Consistency and Invariance Guided Multi-View Hypergraph Learning for Robust Hyperedge Prediction. ACM Trans. Knowl. Discov. Data 20(2): 33:1-33:24 (2026) [4]Shilin Qu, Weiqing Wang, Yuan-Fang Li, Quoc Viet Hung Nguyen, Hongzhi Yin: Scalable and effective negative sample generation for hyperedge prediction. Neural Networks 193: 108034 (2026)

Questions for authors
  1. Existing hyperedge prediction methods do not clearly present the existing problems, which makes it difficult for readers to understand the research motivation of this paper.
  2. The literature review is not comprehensive enough. It does not cover papers published in 2026, and also lacks methods published in 2026 for experimental comparison, which affects the persuasiveness of the paper's experimental results and the innovation of the paper.
  3. The idea of the solution proposed in this paper is relatively outdated and does not meet the publication requirements of top conferences. Conducting node representation learning from the aspects of structure and semantics and adjusting weights accordingly for fusion is a common practice in graph learning papers, lacking novelty.
Reviewer_zJ6POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FAME-HP uses dual-view fuzzy memberships to construct structural, attribute, and cross-view evidence for hyperedge prediction. A candidate-adaptive fusion module combines these signals with a frozen base predictor. Experiments on six hypergraphs evaluate four negative-sampling protocols, with ablations examining the contribution of the proposed components.

Critical strengths
  1. The paper considers an important limitation of hyperedge prediction: different candidate hyperedges may rely on different types of evidence. Separating structural compatibility, attribute compatibility, and cross-view consistency provides a clear motivation for the proposed framework. The method also retains overlapping node affiliations rather than forcing each node into a single community.
  2. The paper examines performance across different datasets and negative-sampling conditions rather than relying on a single aggregate result. The sensitivity analysis also shows how the community dimension and training coefficients affect performance. These experiments help clarify where the method works well and which settings are important for reproducing its results.
Critical weaknesses

1.The methodological novelty of FAME-HP is not sufficiently clear. The framework does not clearly establish which part introduces a fundamentally new modeling mechanism beyond integrating these existing components. The authors should better distinguish the proposed method from existing hyperedge prediction and fuzzy community modeling approaches, and clarify what methodological advance is provided by the candidate-adaptive fusion design. 2. The construction of candidate-level evidence relies on manually specified aggregation weights, but the paper provides limited justification for why these particular combinations are appropriate for hyperedge formation. Moreover, the subsequent adaptive fusion module does not directly explain why the predefined evidence scores are suitable representations of the underlying formation mechanisms. 3.The main comparison uses baseline results reported in previous publications rather than evaluating all methods within a common experimental pipeline. This makes it difficult to determine how much of the reported improvement comes from the proposed method rather than differences in experimental settings. The authors should clarify these differences and provide a more directly comparable evaluation where possible. 4. The candidate-adaptive fusion mechanism is central to the paper, but its contribution is not sufficiently isolated. A comparison against fixed-weight fusion using the same input features would help establish whether candidate-specific weighting is necessary.

Questions for authors

1.What explains the differing results at the same dimension? Was model selection separated from the final evaluation under comparable tuning budgets? 2.Which convergence result applies to the implemented SQN update? If none applies, how will the claim be restricted?

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

This paper proposes FAME-HP for hyperedge prediction in incomplete hypergraphs. Its core concept involves modeling structural information and node attribute semantics separately through dual-view fuzzy overlapping membership, while constructing evidence for each candidate hyperedge based on structure, semantics, and cross-view consistency. Furthermore, the method incorporates 18-dimensional hand-crafted contextual features and utilizes a frozen AllSet-style hypergraph encoder as a base predictor. Subsequently, two MLPs generate evidence correction terms and candidate-adaptive gating weights, respectively, to fuse multi-source evidence and refine the base prediction scores. Experiments were conducted on six real-world hypergraph datasets using four different negative-sampling evaluation protocols, with comparisons made against various existing hyperedge prediction methods. While the approach achieves competitive AUROC and AP scores across several datasets, its core innovation, the consistency within the cross-view membership space, and the independent contribution of the candidate-adaptive fusion mechanism warrant further substantiation.

Critical strengths
  • The paper models three sources of evidence: structural patterns, attribute semantics, and cross-view consistency. This provides a broader view of candidate hyperedge formation than a single unified representation.

  • The model uses structural features such as sub-hyperedge support, prototype similarity, and closure consistency, with a candidate-adaptive gate to weight different evidence sources. It systematically considers various important structural information in hyperedge representation learning.

  • Extensive experiments cover six real-world datasets and four negative sampling protocols. Under the CNS protocol, the model outperformed existing baseline methods in terms of AP across all six datasets.

Critical weaknesses
  • Structural and attribute evidence relies largely on pairwise cosine similarities between node membership vectors and their simple statistics. It remains unclear whether these features capture higher-order compatibility or reliably identify useful formation cues.
  • The evidence branches may learn overlapping or compensatory functions. Without explicit source-specific constraints or further analysis, it is unclear whether each branch captures distinct information, and the claimed interpretability remains unverified.
  • Normalized membership entropy is interpreted as membership uncertainty, an indicator describing the concentration of membership distribution. The paper does not clarify how this measure differs from membership overlap or actual prediction uncertainty.
  • The main baseline uses results reported in existing papers, parameter design and experimental fairness require further supplementation.
  • The paper uses 18 handcrafted structural and prototype-matching features, but their selection rationale is unclear. The separate contributions of NMF-based membership representations and adaptive gating also need to be established.
  • Ablation experiments are insufficient, lacking key comparisons such as simple feature fusion, fixed gating, global gating, context-enhanced base predictors, and node embedding models with matched parameter sizes.
Questions for authors

See Weakness.

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

Related papers

Discussion (0)

Sign in to comment.