Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Deep k-grouping: An Unsupervised Learning Framework for Combinatorial Optimization on Graphs and Hypergraphs

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Deep k-grouping claims that an unsupervised pipeline—one-hot QUBO/PUBO objectives relaxed through softmax, trained by GNNs/HyperGNNs, and annealed with a Gini coefficient term—outperforms SCIP, Tabu, GAP, hMETIS, and KaHyPar on large…

desk verdict Solid engineering, but the headline outperformance claim breaks on the tables: partition comparisons ignore balance, and the runtime test is biased. read the letter →

arxiv 2505.20972 v1 pith:4WPA4MYA submitted 2025-05-27 cs.LG cs.AI

Sen Bai , Chunqi Yang , Xin Bai , Xin Zhang , Zhengang Jiang This is my paper
classification cs.LGcs.AI
keywords combinatorialoptimizationunsupervisedlearninggraphcoloringhypergraphpartitioningQUBOPUBOGinicoefficientneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that unsupervised neural networks can solve k-grouping combinatorial optimization problems—assigning vertices to one of k groups to satisfy coloring or partitioning objectives—at scales where classical exact solvers fail or time out. It introduces one-hot encoded quadratic and polynomial unconstrained binary optimization formulations to express these problems as differentiable losses, then trains graph and hypergraph neural networks to minimize them without any labeled solutions. A Gini-coefficient annealing schedule nudges the continuous softmax outputs toward discrete one-hot assignments while helping the optimizer escape local optima. If the central claim holds, the framework provides a general, GPU-friendly, label-free template for a broad class of NP-hard grouping problems.

What carries the argument

The central object is the one-hot encoded polynomial unconstrained binary optimization (OH-PUBO) formulation, whose variables are one-hot vectors $x_i\in\{0,1\}^k$ and whose terms are Hadamard products such as $x_i\odot x_j$ for edges or $x_{j_1}\odot\cdots\odot x_{j_{|e|}}$ for hyperedges. The quadratic case, OH-QUBO, is written as $\sum_{i,j} Q_{ij}x_i\odot x_j$, with GPU-friendly equivalent $\sum X^T Q \odot X^T$; the polynomial case uses incidence matrices and broadcasting, $\sum Q \odot^{(B)} \mathrm{ColM}(X \odot^{(B)} H +^{(B)} (1-H),1)$. The second load-bearing mechanism is the Gini penalty $\phi(x)=\gamma\sum_i(1-\sum_j x_{ij}^2)$, whose coefficient $\gamma$ is annealed from negative to positive to control the trade-off between exploration and discreteness during unsupervised training.

What would settle it

Run the same pipeline on 100 random graphs with about 200 vertices, compare the final discrete max-cut value to a brute-force optimum, and record the fraction of vertices that fail to converge to a one-hot vector; if the Gini schedule frequently misses the optimum or leaves a non-negligible fraction of vertices fractional, the core annealing claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that k-grouping problems on graphs and hypergraphs can be re-expressed as unconstrained polynomial objectives over one-hot vectors, and that minimizing the softmax relaxation of these objectives with neural networks yields discrete, feasible assignments whose quality matches or beats classical and neural baselines at scale. The paper introduces OH-QUBO for quadratic costs and OH-PUBO for higher-order hyperedge correlations, together with GPU-accelerated formulas for evaluating them. It further introduces a Gini coefficient-based penalty, $\phi(x)=\gamma\sum_i (1-\sum_j x_{ij}^2)$, annealed from negative to positive values, to first smooth the landscape for broad exploration and then enforce discreteness. Reported experiments on graph and hypergraph coloring and partitioning show favorable comparisons against SCIP, Tabu search, GAP, hMETIS, and KaHyPar, especially on large instances.

Load-bearing premise

The approach depends on a single temperature-like schedule—Gini penalty coefficient starting negative and moving positive—to make the network's outputs snap to valid one-hot assignments while escaping local optima, and the paper's ablations cover only a few synthetic cases, so the transfer of that schedule to new datasets is assumed rather than demonstrated.

Editorial extensions

If this is right

  • If the central claim is correct, graph and hypergraph coloring and partitioning instances with tens of thousands of vertices—instances SCIP cannot solve within hours—become solvable in seconds to minutes on a single GPU.
  • The OH-PUBO formulation gives a uniform way to encode any k-grouping objective with hyperedge correlations, potentially extending to other problems such as scheduling, resource allocation, and VLSI partitioning.
  • The negative-to-positive Gini schedule is a simple discretization heuristic that may transfer to other unsupervised combinatorial losses beyond the k-grouping family.
  • Because the framework produces feasible assignments quickly, it could serve as a warm-start generator for exact solvers or refinement heuristics.
  • The GPU-accelerated loss formulas make the approach scalable without specialized hardware beyond standard deep learning infrastructure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the reported superiority is largely a scalability result—on small hypergraphs the paper's own tables show SCIP often reaching the optimal chromatic number while Deep k-grouping is faster but not better, so a fair reading is that the method wins where exact solvers run out of time.
  • Beyond the paper: a natural testable extension is replacing the fixed Gini annealing curve with an adaptive or per-instance schedule, which could reduce sensitivity to the choice of $\gamma$ and training epochs.
  • Beyond the paper: because feasibility is enforced only in the limit of convergence, adding a rounding or projection step after training would be a straightforward extension and could change reported solution values on difficult instances.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Deep k-grouping, an unsupervised learning framework for k-grouping combinatorial optimization problems on graphs and hypergraphs. The authors introduce one-hot encoded quadratic/polynomial unconstrained binary optimization (OH-QUBO/OH-PUBO) formulations, GPU-accelerated tensor implementations of the resulting loss functions, and a Gini-coefficient-based continuous relaxation annealing strategy to encourage discrete, feasible assignments. The framework is evaluated on graph coloring, hypergraph proper coloring, graph partitioning, and hypergraph partitioning against SCIP, Tabu, GAP, hMETIS, and KaHyPar. The paper's central claim is that Deep k-grouping outperforms existing neural network solvers and classical heuristics such as SCIP and Tabu. My assessment is that the methodological formulation is coherent and the GPU cost expressions are a useful contribution, but the empirical evidence does not substantiate the headline claim: the paper's own tables show SCIP achieving better objective values in several graph-coloring, hypergraph-coloring, and partitioning cases, and the partitioning comparisons are not on equal terms because many Deep k-grouping solutions violate the balance constraints while SCIP solutions are perfectly balanced.

Significance. If the empirical claims were supported, the paper would offer a meaningful step toward unified, unsupervised, GPU-accelerated solvers for k-grouping problems, and the OH-PUBO formulation in Eqs. (4)-(5) is genuinely elegant and worth building on. The appendices provide explicit derivations for the tensorized cost functions and the authors honestly disclose one limitation in Sec. 3.1 by acknowledging that SCIP outperforms the method on several small datasets. However, the paper's main advertised contribution is the broad empirical superiority statement in the abstract, and that statement is contradicted by the paper's own experimental tables. The runtime comparison is also not an apples-to-apples measure, and the reported partitioning results do not account for balance violations. The annealed Gini penalty is an interesting idea, but the ablation evidence is limited to synthetic max-cut instances and a single partitioning experiment, with no demonstration that the heuristic transfers across the real-world datasets in Tables 1-3. Because the central load-bearing claim is the empirical result, the paper in its current form cannot be accepted.

major comments (4)
  1. [Abstract and §3.1, Tables 1-2] The abstract's blanket claim that Deep k-grouping outperforms SCIP and Tabu is contradicted by the paper's own experiments. In Table 1, SCIP finds a better chromatic number on UAT (60 vs. 65) within the reported budget and matches or improves on the other small graphs; in Table 2, SCIP finds better chromatic numbers on Primary (27 vs. 35) and Cooking200 (2 vs. 3), and the paper itself states in Sec. 3.1 that 'SCIP outperforms Deep k-grouping and Tabu on small datasets such as BAT, EAT, and UAT.' The stated contribution is empirical superiority, so this self-contradiction invalidates the headline claim and needs to be resolved by rewriting the claims or by substantially strengthening the evidence.
  2. [§3.2, Table 3] The partitioning comparisons are not made on equal terms. All SCIP rows report B1=B2=0, meaning perfectly balanced partitions, while many Deep k-grouping rows have substantial imbalance; for example, EAT k=6 reports Ours C=2,871 with B1=43%, B2=13 versus SCIP C=4,965 with B1=B2=0, and UAT k=6 reports Ours C=2,719 with B1=20%, B2=18 versus SCIP C=11,159 with B1=B2=0. Because Eq. (9) includes the balance penalty β Σ_k (|P_k| - |V|/K)^2, an imbalanced solution is minimizing a different objective, and reporting lower cut values without accounting for balance does not support the claim that Deep k-grouping 'outperforms' SCIP. The comparison should be restricted to feasible, balanced solutions, or the balance constraint should be treated as part of the reported objective rather than as a free parameter.
  3. [§3.3, Fig. 2] The runtime comparison stops SCIP 'immediately when it found a feasible solution,' so Fig. 2 measures time-to-first-feasible, not time to a solution of comparable quality. The conclusion that 'Deep k-grouping achieves better performance on large-scale datasets' is unsupported by this experiment, because SCIP was not given the opportunity to improve its incumbent. A fair comparison would either fix the target solution quality and compare time-to-quality, or fix a time budget and report the best objective value found within that budget. As written, the runtime result conflates feasibility with solution quality and does not substantiate the paper's performance claims.
  4. [§2.3, §3.4] No post-hoc feasibility verification is reported. The method minimizes softmax-relaxed losses, and the Gini penalty in Eq. (12) only encourages discreteness; it does not by itself guarantee that the rounded assignment satisfies all edge or hyperedge constraints. Without a rounding rule and a check that the rounded one-hot assignments yield valid colorings or balanced partitions, the reported objective values (e.g., C in Table 3, chromatic numbers in Tables 1-2) may not correspond to feasible solutions of the same problems solved by SCIP, Tabu, hMETIS, and KaHyPar. This is load-bearing because both coloring and partitioning objectives are only meaningful for feasible discrete assignments. The paper should report constraint-violation counts and the quality of the final rounded solution, and the annealing ablation should be extended to the actual problem families in Tables 1-3 rather than only synthetic max-cut and one k=4 partitioning case.
minor comments (5)
  1. [§3.1] The sentence 'Notably, SCIP's significant performance is achieved through novel objective functions we proposed in Sec. 2.2' is confusing and self-undermining. If SCIP is run on the same objective functions as Deep k-grouping, that should be stated explicitly, but it does not help the comparison; it merely underscores that the proposed formulation is not the source of the claimed empirical advantage.
  2. [Tables 1-2] The tables list two SCIP results per dataset (e.g., 18 (2 s) then 16 (16 s) for BAT) without explaining the reporting convention. The authors should specify whether these are two runs, the first feasible incumbent and the best incumbent within the time limit, or some other convention; the same applies to the single Ours entry per dataset.
  3. [§3.2, Table 3] Table 3 reports no runtime for any partitioning method, but the text in Sec. 3.2 discusses solution spaces and scalability. The claim that Deep k-grouping outperforms the baselines should be accompanied by the time budget for each method so that solution quality can be interpreted in a time-aware manner.
  4. [Appendix D, Eq. (3)] Appendix D correctly notes that Eq. (3) holds only for discrete one-hot vectors and that the softmax relaxation requires separate handling of diagonal and off-diagonal terms as in Eq. (20). This important caveat should appear in the main text where Eq. (3) is first introduced, since the reader might otherwise believe the relaxed loss exactly equals the OH-QUBO objective.
  5. [Figure 1] Figure 1 contains dense and inconsistently rendered labels (e.g., 'grouping Deep −k' and hand-drawn-style mathematical annotations). A cleaner, vectorized figure with consistent notation would substantially improve readability, especially because the figure is the paper's main overview of the proposed framework.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the OH-PUBO losses are direct reformulations of the stated objectives, and the empirical concerns are about experimental fairness, not circular derivation.

full rationale

The paper's mathematical derivation chain consists of Eq. (3), which rewrites the OH-QUBO objective using matrix Hadamard products, and Eq. (5), which rewrites the OH-PUBO objective using the hypergraph incidence matrix. Both are proven in Appendices D and E as algebraic equivalences to the definitions in Eqs. (2) and (4); they are efficient implementations of the stated objectives, not independent results predicted from fitted inputs. The problem-specific cost functions in Section 2.2 are formulated directly from the coloring and partitioning requirements, and the Gini annealing term in Eq. (12) is presented as a heuristic inspired by the external CRA-PI-GNN work, with ablations evaluating solution quality rather than any held-out prediction. The experimental comparison with SCIP raises legitimate fairness and correctness concerns, such as stopping SCIP at the first feasible solution and reporting cut values for Deep k-grouping solutions that violate the balance constraint, but those are not circularity: they concern whether the empirical claim is valid, not whether the derivation reduces to its own inputs. The paper also explicitly concedes that SCIP outperforms Deep k-grouping on BAT, EAT, and UAT, contradicting the abstract's blanket claim; again, this is an evidence-quality issue rather than a circular-logic issue. No load-bearing self-citation or author-imported uniqueness argument appears; the cited prior work is external and not used to forbid alternatives. Therefore, under the specified circularity criteria, no step in the paper is circular by construction, and the honest finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The framework relies on standard one-hot encoding and penalty methods; no new physical entities are introduced. The main free parameters are the penalty and annealing schedules, which are hand-tuned. The Gini annealing is an ad hoc variant of the existing CRA idea.

free parameters (5)
  • λ1, λ2 penalty coefficients = Initially 0, gradually increased
    Used in Eqs. 6-8 to enforce coloring constraints; the schedule is hand-chosen and not reported per dataset.
  • γ Gini annealing strength = Negative to positive (e.g., -2.5 to >0 in ablation)
    Eq. 12; the annealing schedule is tuned by hand for each experiment in Sec. 3.4; no automatic selection or sensitivity analysis.
  • α/β ratio = Fixed value, not specified
    In Eqs. 9-10; the ratio is a hyperparameter, but the actual value or range is never given.
  • Learning rate η = 1e-4
    Set in Secs. 3.1 and 3.2; no schedule or sensitivity analysis.
  • Architecture hyperparameters = Not reported
    GNN depth (2-3 layers for coloring, 4-8 for partitioning) and hidden dimensions are mentioned vaguely; exact values are missing.
assumptions (3)
  • domain assumption Softmax relaxation of one-hot variables preserves the ordering of the discrete OH-QUBO/OH-PUBO objective well enough that gradient descent finds good discrete solutions
    Invoked throughout Sec. 2.3; no theory or guarantee is provided.
  • domain assumption GNN/HyperGNN can be trained from random node features to minimize the combinatorial loss
    Used in Eq. 11 and experiments; no random baseline is compared.
  • ad hoc to paper The Gini penalty 1 - sum_j x_ij^2 decreases as the softmax output becomes more peaked, and annealing γ from negative to positive avoids local optima
    Sec. 2.4 justifies this by analogy to CRA-PI-GNN, not by derivation or systematic study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep k-grouping: An Unsupervised Learning Framework for Combinatorial Optimization on Graphs and Hypergraphs." pith.science (2026). https://pith.science/paper/4WPA4MYA

@misc{pith2026250520972,
  author       = {Pith},
  title        = {Pith review of: Deep k-grouping: An Unsupervised Learning Framework for Combinatorial Optimization on Graphs and Hypergraphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4WPA4MYA}},
  note         = {Machine review of arXiv:2505.20972}
}
abstract

Along with AI computing shining in scientific discovery, its potential in the combinatorial optimization (CO) domain has also emerged in recent years. Yet, existing unsupervised neural network solvers struggle to solve $k$-grouping problems (e.g., coloring, partitioning) on large-scale graphs and hypergraphs, due to limited computational frameworks. In this work, we propose Deep $k$-grouping, an unsupervised learning-based CO framework. Specifically, we contribute: Novel one-hot encoded polynomial unconstrained binary optimization (OH-PUBO), a formulation for modeling k-grouping problems on graphs and hypergraphs (e.g., graph/hypergraph coloring and partitioning); GPU-accelerated algorithms for large-scale k-grouping CO problems. Deep $k$-grouping employs the relaxation of large-scale OH-PUBO objectives as differentiable loss functions and trains to optimize them in an unsupervised manner. To ensure scalability, it leverages GPU-accelerated algorithms to unify the training pipeline; A Gini coefficient-based continuous relaxation annealing strategy to enforce discreteness of solutions while preventing convergence to local optima. Experimental results demonstrate that Deep $k$-grouping outperforms existing neural network solvers and classical heuristics such as SCIP and Tabu.

Figures

Figures reproduced from arXiv: 2505.20972 by the authors.

Figure 1
Figure 1. Overview of unsupervised neural network-based CO frameworks, including QUBO-based neural network [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the run￾time between SCIP and Deep k￾grouping. We evaluate the performance of Deep k-grouping on graph/hypergraph partitioning as depicted in Tab. 3. Best performed GAT [36] with 2 or 3 layers is applied to construct Deep k-grouping for graph partitioning. For hypergraph partitioning, Deep k-grouping achieves best solutions with 4 ∼ 8-layer HGNN+ [33] incorporating FC layers. α β is assigned a fixed va… view at source ↗
Figure 5
Figure 5. Quality and discreteness of solutions of the graph partition￾ing problem with or without Gini coefficient-based annealing strategy. We further evaluated the Gini coefficient-based annealing strategy for the graph partitioning problem (k = 4). The experiments were conducted on a graph comprising 5,000 vertices and 5,000 edges. The penalty strength γ was set to −0.25 and reached 0 after 1000 epochs and continued to in… view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: On the proof of Eq. 3. : v1 v3 v2 v4 e3 e2 e1 0.70.3 0.10.9 0.20.8 0.01.0 1 1 0 1 v1 v2 v3 v4 1 0 1 0 1 e1 e2 e3 0.70.7 0 0.1 0 0.1 0 0.20.2 0 0.0 0.0 0.30.3 0 0.9 0 0.9 0 0.80.8 0 1.0 1.0 0 1 1 0.70.7 1 0.1 1 0.1 1 0.20.2 1 0.0 0.0 0.30.3 1 0.9 1 0.9 1 0.80.8 1 1.0 1.…
Figure 7
Figure 7. Figure 7: A toy example for Eq. 5. where xi ∈ {0, 1} 2 are 2-dimensional one-hot vectors, and the matrix [x1, x2, ..., x|V | ] is denoted by X ∈ R |V |×2 . The OH-QUBO matrix Q=    2 if (i, j) ∈ E −2 if i = j 0 otherwise . D Proof of GPU-accelerated OH-QUBO cost function in E…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [1]

    Gnn&gbdt-guided fast optimizing framework for large-scale integer programming

    Huigen Ye, Hua Xu, Hongyan Wang, Chengming Wang, and Yu Jiang. Gnn&gbdt-guided fast optimizing framework for large-scale integer programming. In International conference on machine learning (ICML), pages 39864–39878. PMLR, 2023

    work page 2023

  2. [2]

    Neural network-based dimensionality reduction for large-scale binary optimization with millions of variables

    Ye Tian, Luchen Wang, Shangshang Yang, Jinliang Ding, Yaochu Jin, and Xingyi Zhang. Neural network-based dimensionality reduction for large-scale binary optimization with millions of variables. IEEE Transactions on Evolutionary Computation, 2024

    work page 2024

  3. [3]

    On the power of small-size graph neural networks for linear programming

    Qian Li, Tian Ding, Linxin Yang, Minghui Ouyang, Qingjiang Shi, and Ruoyu Sun. On the power of small-size graph neural networks for linear programming. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 9 A PREPRINT - SEPTEMBER 1, 2025

    work page 2024

  4. [4]

    Combinatorial optimization and reasoning with graph neural networks

    Quentin Cappart, Didier Chételat, Elias B Khalil, Andrea Lodi, Christopher Morris, and Petar Veli ˇckovi´c. Combinatorial optimization and reasoning with graph neural networks. Journal of Machine Learning Research, 24(130):1–61, 2023

    work page 2023

  5. [5]

    Combinatorial optimization with physics-inspired graph neural networks

    Martin JA Schuetz, J Kyle Brubaker, and Helmut G Katzgraber. Combinatorial optimization with physics-inspired graph neural networks. Nature Machine Intelligence, 4(4):367–377, 2022

    2022

  6. [6]

    A deep learning framework for graph partitioning

    Azade Nazi, Will Hang, Anna Goldie, Sujith Ravi, and Azalia Mirhoseini. A deep learning framework for graph partitioning. International Conference on Learning Representations (ICLR), 2019

    work page 2019

  7. [7]

    Graph clustering with graph neural networks

    Anton Tsitsulin, John Palowitch, Bryan Perozzi, and Emmanuel Müller. Graph clustering with graph neural networks. Journal of Machine Learning Research, 24(127):1–21, 2023

    work page 2023

  8. [8]

    Controlling continuous relaxation for combinatorial optimization.Advances in Neural Information Processing Systems (NeurIPS), 37:47189–47216, 2024

    Yuma Ichikawa. Controlling continuous relaxation for combinatorial optimization.Advances in Neural Information Processing Systems (NeurIPS), 37:47189–47216, 2024

    work page 2024

Show all 43 references
  1. [9]

    Learning to solve quadratic unconstrained binary optimization in a classification way

    Ming Chen, Jie Chun, Shang Xiang, Luona Wei, Yonghao Du, Qian Wan, Yuning Chen, and Yingwu Chen. Learning to solve quadratic unconstrained binary optimization in a classification way. Advances in Neural Information Processing Systems (NeurIPS), 37:114478–114509, 2024

    2024

  2. [10]

    Neural graph matching network: Learning lawler’s quadratic assignment problem with extension to hypergraph and multiple-graph matching

    Runzhong Wang, Junchi Yan, and Xiaokang Yang. Neural graph matching network: Learning lawler’s quadratic assignment problem with extension to hypergraph and multiple-graph matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5261–5279, 2021

    2021

  3. [11]

    Quantum bridge analytics i: a tutorial on formulating and using qubo models

    Fred Glover, Gary Kochenberger, Rick Hennig, and Yu Du. Quantum bridge analytics i: a tutorial on formulating and using qubo models. Annals of Operations Research, 314(1):141–183, 2022

    2022

  4. [12]

    Colorrl: reinforced coloring for end-to-end instance segmentation

    Tran Anh Tuan, Nguyen Tuan Khoa, Tran Minh Quan, and Won-Ki Jeong. Colorrl: reinforced coloring for end-to-end instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16727–16736, 2021

    2021

  5. [13]

    A graph coloring approach for image segmentation

    D Gómez, Javier Montero, Javier Yáñez, and C Poidomani. A graph coloring approach for image segmentation. Omega, 35(2):173–183, 2007

    2007

  6. [14]

    Toward quantum gate-model heuristics for real-world planning problems

    Tobias Stollenwerk, Stuart Hadfield, and Zhihui Wang. Toward quantum gate-model heuristics for real-world planning problems. IEEE Transactions on Quantum Engineering, 1:1–16, 2020

    2020

  7. [15]

    Towards sustainable industry 4.0: A green real-time iiot multitask scheduling architecture for distributed 3d printing services

    Lamis R Darwish, Mohamed T El-Wakad, and Mahmoud M Farag. Towards sustainable industry 4.0: A green real-time iiot multitask scheduling architecture for distributed 3d printing services. Journal of Manufacturing Systems, 61:196–209, 2021

    2021

  8. [16]

    A study on course timetable scheduling using graph coloring approach

    Runa Ganguli and Siddhartha Roy. A study on course timetable scheduling using graph coloring approach. International Journal of Computational and Applied Mathematics, 12(2):469–485, 2017

    2017

  9. [17]

    Coloring-based channel allocation for multiple coexisting wireless body area networks: A game-theoretic approach

    Kai-Ju Wu, Y-W Peter Hong, and Jang-Ping Sheu. Coloring-based channel allocation for multiple coexisting wireless body area networks: A game-theoretic approach. IEEE Transactions on Mobile Computing, 21(1):63–75, 2020

    2020

  10. [18]

    Hypergraph- based interference avoidance resource management in customer-centric communication for intelligent cyber- physical transportation systems

    Jie Huang, Shilong Zhang, Fan Yang, Tao Yu, LV Narasimha Prasad, Manisha Guduri, and Keping Yu. Hypergraph- based interference avoidance resource management in customer-centric communication for intelligent cyber- physical transportation systems. IEEE Transactions on Consumer ...

    2023

  11. [19]

    Uav-assisted emergency communica- tions in social iot: A dynamic hypergraph coloring approach

    Bowen Wang, Yanjing Sun, Zhi Sun, Long D Nguyen, and Trung Q Duong. Uav-assisted emergency communica- tions in social iot: A dynamic hypergraph coloring approach. IEEE Internet of Things Journal, 7(8):7663–7677, 2020

    2020

  12. [20]

    A grover search-based algorithm for the list coloring problem

    Sayan Mukherjee. A grover search-based algorithm for the list coloring problem. IEEE Transactions on Quantum Engineering, 3:1–8, 2022

    2022

  13. [21]

    Multilevel hypergraph partitioning: Applica- tion in vlsi domain

    George Karypis, Rajat Aggarwal, Vipin Kumar, and Shashi Shekhar. Multilevel hypergraph partitioning: Applica- tion in vlsi domain. In Proceedings of the 34th annual Design Automation Conference (DAC), pages 526–529, 1997

    1997

  14. [22]

    Mapart: an efficient multi-fpga system-aware hypergraph partitioning framework

    Benzheng Li, Shunyang Bi, Hailong You, Zhongdong Qi, Guangxin Guo, Richard Sun, and Yuming Zhang. Mapart: an efficient multi-fpga system-aware hypergraph partitioning framework. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024

    2024

  15. [23]

    Stochastic shadow detection using a hypergraph partitioning approach

    Vitor Gomes, Pablo Barcellos, and Jacob Scharcanski. Stochastic shadow detection using a hypergraph partitioning approach. Pattern Recognition, 63:30–44, 2017. 10 A PREPRINT - SEPTEMBER 1, 2025

    2017

  16. [24]

    A novel adaptive hypergraph neural network for enhancing medical image segmentation

    Shurong Chai, Rahul K Jain, Shaocong Mo, Jiaqing Liu, Yulin Yang, Yinhao Li, Tomoko Tateyama, Lanfen Lin, and Yen-Wei Chen. A novel adaptive hypergraph neural network for enhancing medical image segmentation. In International Conference on Medical Image Computing and Computer-...

    2024

  17. [25]

    Social hash partitioner: a scalable distributed hypergraph partitioner

    Igor Kabiljo, Brian Karrer, Mayank Pundir, Sergey Pupyrev, and Alon Shalita. Social hash partitioner: a scalable distributed hypergraph partitioner. Proceedings of the VLDB Endowment, 10(11):1418–1429, 2017

    2017

  18. [26]

    Hepart: A balanced hypergraph partitioning algorithm for big data applications

    Wenyin Yang, Guojun Wang, Kim-Kwang Raymond Choo, and Shuhong Chen. Hepart: A balanced hypergraph partitioning algorithm for big data applications. Future Generation Computer Systems, 83:250–268, 2018

    2018

  19. [27]

    Automated distribution of quantum circuits via hypergraph partitioning

    Pablo Andres-Martinez and Chris Heunen. Automated distribution of quantum circuits via hypergraph partitioning. Physical Review A, 100(3):032308, 2019

    2019

  20. [28]

    Hyper-optimized tensor network contraction

    Johnnie Gray and Stefanos Kourtis. Hyper-optimized tensor network contraction. Quantum, 5:410, 2021

    2021

  21. [29]

    Regularizing neural networks by penalizing confident output distributions

    Gabriel Pereyra, George Tucker, Jan Chorowski, Łukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. arXiv preprint arXiv:1701.06548, 2017

    2017 arXiv

  22. [30]

    PySCIPOpt: Mathematical programming in python with the SCIP optimization suite

    Stephen Maher, Matthias Miltenberger, João Pedro Pedroso, Daniel Rehfeldt, Robert Schwarz, and Felipe Serrano. PySCIPOpt: Mathematical programming in python with the SCIP optimization suite. In Mathematical Software – ICMS 2016, pages 301–307. Springer International Publishing, 2016

    2016

  23. [31]

    Tabu search

    Fred Glover and Manuel Laguna. Tabu search. Springer, 1998

    1998

  24. [32]

    Scalable high-quality hypergraph partitioning

    Lars Gottesbüren, Tobias Heuer, Nikolai Maas, Peter Sanders, and Sebastian Schlag. Scalable high-quality hypergraph partitioning. ACM Transactions on Algorithms, 20(1):1–54, 2024

    2024

  25. [33]

    Hgnn+: General hypergraph neural networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181–3199, 2022

    Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181–3199, 2022

    2022

  26. [34]

    Hypergraph neural networks

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks. In the AAAI Conference on Artificial Intelligence (AAAI), volume 33, pages 3558–3565, 2019

    2019

  27. [35]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017

    2017

  28. [36]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations (ICLR), 2018

    2018

  29. [37]

    Contrastive learning network for unsupervised graph matching

    Yu Xie, Lianhang Luo, Tianpei Cao, Bin Yu, and AK Qin. Contrastive learning network for unsupervised graph matching. IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  30. [38]

    Improving graph matching with positional reconstruction encoder-decoder network

    Yixiao Zhou, Ruiqi Jia, Hongxiang Lin, Hefeng Quan, Yumeng Zhao, and Xiaoqing Lyu. Improving graph matching with positional reconstruction encoder-decoder network. Advances in Neural Information Processing Systems (NeurIPS), 36:34557–34569, 2023

    2023

  31. [39]

    Discrete cycle-consistency based unsupervised deep graph matching

    Siddharth Tourani, Muhammad Haris Khan, Carsten Rother, and Bogdan Savchynskyy. Discrete cycle-consistency based unsupervised deep graph matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 38, pages 5252–5260, 2024

    2024

  32. [40]

    Specpart: A supervised spectral framework for hypergraph partitioning solution improvement

    Ismail Bustany, Andrew B Kahng, Ioannis Koutis, Bodhisatta Pramanik, and Zhiang Wang. Specpart: A supervised spectral framework for hypergraph partitioning solution improvement. In the 41st IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1–9, 2022

    2022

  33. [41]

    K-specpart: Supervised embedding algorithms and cut overlay for improved hypergraph partitioning

    Ismail Bustany, Andrew B Kahng, Ioannis Koutis, Bodhisatta Pramanik, and Zhiang Wang. K-specpart: Supervised embedding algorithms and cut overlay for improved hypergraph partitioning. IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems, 2023

    2023

  34. [42]

    Hypergraph partitioning with embeddings

    Justin Sybrandt, Ruslan Shaydulin, and Ilya Safro. Hypergraph partitioning with embeddings. IEEE Transactions on Knowledge & Data Engineering, 34(06):2771–2782, 2022. 11 A PREPRINT - SEPTEMBER 1, 2025 A Related works PI-GNN [5] introduced the QUBO formulation from quantum comp...

    2022

  35. [43]

    vertices belonging to the independent set and adding a penalty when two vertices are connected by an edge. IP Form. Given a graph G = (V, E), the MIS problem can be formulated as the following IP form: min − X i∈V xi (13) s.t. x ixj = 0 for all ( i, j) ∈ E (14) xi ∈ {0, 1} (15...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.