acceptodds
Withdrawn from ICLR 2027

One Model, Three Levels: Multi-Task Learning on Hypergraphs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Developing a single model for multiple tasks has long been a goal in artificial intelligence. In hypergraph learning, existing work has explored using a single model to handle node-level and hyperedge-level tasks, but has not incorporated hypergraph-level tasks into the same model. Therefore, we consider the problem of handling node-level, hyperedge-level, and hypergraph-level tasks within a single model, which poses two challenges. First, these tasks involve prediction targets at different structural levels and differ in their input and prediction formats. Second, their optimization objectives may interfere with one another, so improving performance on one task may degrade performance on another. To address these challenges, we propose a model for multi-task learning on hypergraphs. To unify the input and prediction formats across the three tasks, the proposed model uses query-marked subhypergraphs as a common input representation and applies a common query readout followed by prototype matching to produce predictions in a common format. To mitigate cross-task interference, it combines a shared hypergraph encoder with task-conditioned adapters, enabling the model to capture shared higher-order structural patterns while preserving task-specific information. Experimental results demonstrate that the proposed model performs comparably to the baselines across node-, hyperedge-, and hypergraph-level tasks while mitigating cross-task interference. The code is available at: https://anonymous.4open.science/r/HyperOTri-7EFB.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
hypergraph neural networks, multi-task learning

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 2 reviews, above 9% of the others.
Reviewer_wEpJOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents HyperOTri, a single model for node classification, hyperedge prediction, and hypergraph classification. The goal is clear. The model performs better than Naive MTL on several tasks, but its gains over single-task training are mixed.

I have several concerns. The paper does not explain how its KK-hop neighborhood preserves distance information in distance-based hypergraphs. Section3.2 and the appendix also describe different procedures for constructing subhypergraphs. Although Equation(8) defines the joint loss, the benefit of adding other tasks to a primary task remains unclear. Finally, the paper does not explain why it sets the adapter scale α\alpha to 0.1 or test how sensitive the results are to this choice. These issues weaken the evidence for the method's effectiveness.

Critical strengths

1.The method is clearly designed. Its query-marked subhypergraphs, shared encoder, task-conditioned adapters, and unified prediction module form a coherent framework for the three tasks.

2.The experimental evaluation is relatively comprehensive. It includes multiple datasets, comparisons with single-task and multitask baselines, component ablations, and analyses of several hyperparameters.

Critical weaknesses
  1. The definition and implementation of the core input construction are inconsistent. Section 3.2 defines the local input as a node-induced subhypergraph. Appendix B.4 instead describes collecting incident hyperedges hop by hop, with at most LL hyperedges retained. These procedures can produce different input structures. Subhypergraph construction is central to the proposed framework. Without a precise account of the algorithm and sampling rules used in the experiments, the method is difficult to reproduce. It is also unclear whether the reported gains result from the stated definition or a particular sampling procedure.

  2. The experiments do not sufficiently establish the value of joint learning across three tasks. In Table 1, HyperOTri performs worse than single-task training in four of nine results. Under the three-task setting in Figure 4, node-classification accuracy drops by 1.70 percentage points and hypergraph-classification accuracy drops by 6.70 percentage points. Although the model outperforms Naive MTL, this comparison does not establish consistent benefits for the target tasks. When performance on one task is the primary goal, the current evidence does not sufficiently justify using the proposed method. This substantially affects my assessment of the paper’s central contribution and experimental support.

Questions for authors
  1. Neighborhood definition. Section 3.2 treats any two nodes in the same hyperedge as one hop apart. Is KK-hop distance intended to describe incidence-based connectivity only? For distance-constructed hypergraphs such as NTU2012, does the subhypergraph construction use any information about distances between nodes within a hyperedge?

  2. Subhypergraph construction. Which rule is implemented: retaining all hyperedges fully contained in the KK-hop node set, as stated in Section 3.2, or collecting incident hyperedges at each expansion step, as stated in Appendix B.4? These rules can produce different subhypergraphs. How is the limit of LL hyperedges applied?

  3. Effect of joint training. Equation (8) assigns equal weight to the three task losses. Using the same architecture and evaluation protocol, what is the performance of each target task when trained alone, with each auxiliary task, and with both auxiliary tasks? Please report the changes relative to target-task-only training so that positive and negative transfer can be assessed directly.

  4. Strength of the empirical gains. Several improvements in Table 1 are small, and some results are below the single-task baseline. Which gains over the strongest applicable baseline are consistent across the five runs, and what is the main empirical advantage supported by these results?

  5. Choice of α\alpha. How was the residual scale α=0.1\alpha=0.1 selected? How sensitive are the three task results to α\alpha, including α=0\alpha=0, which removes the adapter contribution?

Reviewer_fwT7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multitask learning on hypergraphs and proposes HyperOTri to jointly address node classification, hyperedge prediction, and hypergraph classification. HyperOTri consists of three stages: unified task input formulation, task-aware representation learning, and unified prediction. Specifically, it first reformulates different tasks as query-marked subhypergraphs, then learns task-aware representations using a shared HGNN with task-conditioned adapters. Finally, a shared query readout and prototype matching are used to generate predictions under a unified classification loss. Experiments on seven datasets across three tasks show that HyperOTri generally outperforms the evaluated multitask baselines.

Critical strengths

S1. The overall design is reasonable. Query-based inputs and task-conditioned adapters provide a coherent way to handle three levels of hypergraph tasks within a unified model.

S2. The paper is generally clear and easy to follow. The figures and equations explain the main pipeline well.

Critical weaknesses

W1. The methodological contribution appears incremental. Several components of HyperOTri have been explored in prior graph multitask learning methods. Query-marked subhypergraphs follow task reformulation strategies explored in All in One [1] and OFA [2], while the use of a shared GNN with task-specific adapters is also seen in G-Merging [3]. It remains unclear what technical advances HyperOTri offers beyond adapting and combining these existing mechanisms for hypergraphs.

W2. The motivation needs clearer positioning against existing foundation models. The introduction lacks a discussion of hypergraph foundation models such as Hyper-FM [4]. It remains unclear what gaps in unified task formulation and cross-task interference motivate HyperOTri beyond existing work.

W3. The experimental coverage is limited. Each task is evaluated on only three relatively small datasets, and two of the three hypergraph classification datasets are synthetic.

W4. The released code is inconsistent with the method description. The query marker is not appended to the HGNN input features in the released implementation, as described in Section 3.2.

W5. The citation formatting needs correction. Citations in parentheses should use \citep{}, while \citet{} should be used when the citations are part of the sentence.

[1] Sun et al. All in One: Multi-Task Prompting for Graph Neural Networks. KDD, 2023.

[2] Liu et al. One for All: Towards Training One Graph Model for All Classification Tasks. ICLR, 2024.

[3] Chen et al. G-Merging: Graph Models Merging for Parameter-Efficient Multi-Task Knowledge Consolidation. ICLR, 2026.

[4] Gao et al. Hypergraph Foundation Model. IEEE TPAMI, 2026.

Questions for authors

Q1. What are the key contributions of HyperOTri beyond adapting and combining existing graph multitask learning mechanisms?

Q2. Could the authors discuss whether the two identified challenges have been well studied in existing methods, such as Hyper-FM, and provide a clearer analysis of the remaining challenges in multitask learning on hypergraphs?

Q3. Can the authors provide results on additional and larger real-world datasets to better demonstrate the generalizability and scalability of HyperOTri?

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

Related papers

Discussion (0)

Sign in to comment.