pith. machine review for the scientific record. sign in

arxiv: 2605.09382 · v1 · submitted 2026-05-10 · 💻 cs.LG · cs.CV· cs.DS· math.OC

Recognition: 2 theorem links

· Lean Theorem

Learning-Augmented Scalable Linear Assignment Problem Optimization via Neural Dual Warm-Starts

Ilay Yavlovich , Jad Agbaria , Muhamed Mhamed , Jose Yallouz , Nir Weinberger
Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:17 UTC · model grok-4.3

classification 💻 cs.LG cs.CVcs.DSmath.OC
keywords linear assignment problemwarm-startdual variablesneural networkcombinatorial optimizationLAPJVscalabilitylearning-augmented algorithms
0
0 comments X

The pith

A neural network predicts dual variables to warm-start exact solvers for the linear assignment problem, achieving over 2x speedups while guaranteeing optimality.

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

The paper shows how a lightweight neural architecture can predict dual variables to initialize classical exact solvers like LAPJV for the linear assignment problem. This hybrid approach speeds up computation on large instances without sacrificing the exact optimality that pure learning methods often lose. It matters for applications in tracking, vision, and logistics where assignment problems arise at scale and must be solved precisely. The design uses a row-independent network to avoid memory issues and a simple duality trick to keep solutions feasible.

Core claim

Predicting dual variables with RowDualNet allows the Jonker-Volgenant algorithm to find optimal assignments with significantly reduced search effort, delivering over 2× speedups on synthetic distributions and 1.25× to 1.5× improvements on real-world MOT tracking and LPT transportation datasets, all while preserving full optimality through a fallback mechanism and enabling scaling to N=16,384.

What carries the argument

RowDualNet, a row-independent neural network that predicts dual variables for each assignment row separately, together with the Min-Trick that constructs feasible duals from LP duality to warm-start the solver.

If this is right

  • Classical exact solvers can be made faster on large LAP instances without any loss of optimality.
  • The approach scales to problems with 16,384 variables due to linear memory usage.
  • Robust zero-shot performance is achieved on real-world datasets from tracking and transportation.
  • Worst-case runtime guarantees are maintained via the fallback when predictions are poor.
  • Search effort in the solver is reduced on challenging synthetic distributions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This method could extend to accelerating other primal-dual based combinatorial optimization algorithms.
  • Hybrid learning-exact approaches like this may become standard for balancing speed and guarantees in optimization.
  • Further gains might come from training the network on more diverse distributions or integrating it with adaptive solvers.
  • Applications in dynamic or online assignment settings could benefit from the fast warm-starting.

Load-bearing premise

That the dual variable predictions are accurate enough on new, unseen problem distributions to reduce the solver's internal search steps enough to offset the cost of running the network.

What would settle it

A set of large-scale LAP instances from a new distribution where the total runtime of the neural warm-start method equals or exceeds the runtime of the unassisted LAPJV solver.

Figures

Figures reproduced from arXiv: 2605.09382 by Ilay Yavlovich, Jad Agbaria, Jose Yallouz, Muhamed Mhamed, Nir Weinberger.

Figure 1
Figure 1. Figure 1: Pipeline for Neural Dual Warm-Starts. We extract row-centric features, predict row potentials uˆ via RowDualNet, and construct feasible column potentials vˆ using the Min-Trick. This valid dual pair seeds the LAPJV solver. A fallback mechanism (red path) ensures robustness by reverting to a cold start if the prediction quality is low. maintaining O(N · D) complexity, enabling scalability to massive instanc… view at source ↗
Figure 2
Figure 2. Figure 2: Scalability on Dense Model. End-to-end speedup factor (Tbaseline/Tours) versus matrix size N on dense matrices (Mean of Ratios ± 95% CI). The solid lines indicate the mean speedup, while the shaded regions represent the 95% CI. Our Neural Warm-Start framework consistently outperforms both Cold￾Start LAP and SciPy when N ≥ 1,024, reaching a peak speedup of ≈ 2.5×. The widening shaded area for LAP (green) re… view at source ↗
Figure 4
Figure 4. Figure 4: Generalization to Urban Transportation Networks. End-to-end speedup factor on cost matrices derived from Open￾StreetMap (OSM) for N = 10,000 locations in major global metropolises. Our method achieves consistent acceleration (≈ 1.2× – 1.8×) across diverse urban topologies, demonstrating ro￾bustness to real-world road network structures without fine-tuning. The error bars indicate the 95% CI over 10 random … view at source ↗
Figure 6
Figure 6. Figure 6: Comparison with Dinitz et al. (2021). Speedup factor relative to Cold-Start baselines (average of each problem size). On the Block-Structured Model dataset, the static “Learned Median” baseline from (Dinitz et al., 2021) (dashed lines) fails to scale, dropping below 1.0× for all instances. In contrast, Neural Dual Warm-Starting (solid lines) achieves superior acceleration, reach￾ing up to ≈ 3.5× against LA… view at source ↗
Figure 8
Figure 8. Figure 8: RowDualNet Architecture. The model processes row features through a residual MLP to produce an initial dual estimate uˆpre. The Sparse Refine Mechanism (center) then injects global context by inspecting the top-K most competitive columns (lowest reduced cost) for each agent, allowing the network to adjust bids based on contention before the final output. at exact ties, standard subgradient handling in auto… view at source ↗
Figure 9
Figure 9. Figure 9: Solver Work Analysis for Block-Structured Model. (Left) The average percentage of rows matched during the initial greedy phase. Standard Cold-Start heuristics (SciPy, Red) consistently resolve only ≈ 26% of assignments. In contrast, our Neural Warm-Start (Green) matches 75.2% of agents immediately. (Right) This superior initialization drastically reduces the number of “free rows” requiring expensive augmen… view at source ↗
Figure 10
Figure 10. Figure 10: Solver Work Analysis for Dense Model. (Left) The average percentage of rows matched during the initial greedy phase. Standard Cold-Start heuristics (SciPy, Red) consistently resolve only ≈ 27% of assignments. In contrast, our Neural Warm-Start (Green) matches 76.4% of agents immediately. (Right) This superior initialization drastically reduces the number of “free rows” requiring expensive augmenting path … view at source ↗
Figure 11
Figure 11. Figure 11: Ablation Study on Dense Model. Comparison of warm-start strategies (Mean of Ratios) relative to the Cold-Start baseline (1.0×). While Linear Regression (pink) offers marginal gains for small instances, its performance degrades as N increases, confirming that simple models cannot capture the complexity of large, dense assignment problems. Heuristics like Random and Row Mean (yellow/green) consistently perf… view at source ↗
Figure 12
Figure 12. Figure 12: Ablation Study on Block-Structured Model. Speedup factors on Block-Structured Model instances (Mean of Ratios). The Linear Regression baseline (pink) fails to generalize as problem size grows, eventually falling below the cold-start baseline (1.0×) at large N. In contrast, Neural Dual Warm-Starts (orange) successfully exploits the latent group structure, achieving superior speedups of up to ≈ 3.5× against… view at source ↗
Figure 13
Figure 13. Figure 13: Overhead of a Failed Warm-Start. The relative time penalty incurred when the fallback mechanism is triggered. As the problem size N grows, the fixed cost of neural inference and safety checking (O(N 2 )) becomes negligible compared to the solver’s cubic complexity (O(N 3 )). I. Extended Related Work This section provides additional context on related approaches that were summarized in Sec. 7. I.1. Classic… view at source ↗
read the original abstract

The Linear Assignment Problem (LAP) is a fundamental combinatorial optimization task with applications ranging from computer vision to logistics. Classical exact solvers such as the Hungarian and Jonker-Volgenant (LAPJV) algorithms guarantee optimality, but their cubic time complexity $\mathcal{O}(N^{3})$ becomes a bottleneck for large-scale instances. Recent learning-based approaches aim to replace these solvers with neural models, often sacrificing exactness or failing to scale due to memory constraints. We propose a learning-augmented framework that accelerates exact assignment solvers while maintaining optimality and worst-case guarantees. Our method predicts dual variables to warm-start a classical solver, with a fallback that prevents asymptotic runtime degradation when the learned advice is unreliable. We introduce RowDualNet, a lightweight row-independent architecture that avoids the $\mathcal{O}(N^{2})$ memory bottleneck of graph-based models, enabling neural warm-starting at large scale ($N=16{,}384$). Feasibility is ensured via a constructive mechanism based on LP duality (namely, the Min-Trick), eliminating costly iterative projection. Empirically, our approach reduces the search effort of LAPJV and achieves over $2{\times}$ speedups on challenging synthetic distributions, in addition to improving over $1.25{\times}$ and $1.5{\times}$ on real-world tracking (MOT) and transportation (LPT) datasets, respectively, while strictly maintaining full optimality, effectively yielding a robust zero-shot generalization to real-world tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a learning-augmented framework for the Linear Assignment Problem that uses a lightweight, row-independent neural network (RowDualNet) to predict dual variables for warm-starting the LAPJV solver. A constructive Min-Trick adjustment ensures feasibility of the predicted duals, while a classical fallback mechanism guarantees optimality and prevents asymptotic slowdown. The approach is claimed to deliver >2× speedups on challenging synthetic distributions and 1.25–1.5× improvements on real-world MOT tracking and LPT transportation instances while preserving exact optimality and enabling zero-shot generalization.

Significance. If the empirical claims hold under closer scrutiny, the work provides a pragmatic bridge between learned heuristics and exact combinatorial solvers, preserving worst-case guarantees while scaling to N=16k instances. The row-independent architecture and non-iterative Min-Trick are concrete strengths that address memory and projection costs of prior graph-based methods. The fallback design is a positive example of robust learning-augmented optimization.

major comments (2)
  1. [Experiments] Experiments section: aggregate speedups are reported for MOT and LPT, yet the manuscript provides no statistics on fallback invocation frequency, post-Min-Trick duality gaps, or the correlation between per-instance prediction error and observed runtime reduction. Without these, it is impossible to determine whether RowDualNet consistently reduces LAPJV search effort on out-of-distribution data or whether gains are driven by a minority of instances.
  2. [Method] Method and § on RowDualNet: because the network predicts duals independently per row, the paper should quantify how well the resulting (Min-Trick-adjusted) duals approximate the optimal dual solution on real-world instances; an ablation relating prediction error to search-tree pruning would directly test the central assumption that learned advice reliably accelerates the solver before fallback.
minor comments (1)
  1. [Abstract] The abstract and introduction would benefit from a brief statement of the precise form of the fallback (e.g., when it triggers and what solver state is reused).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's significance, particularly the row-independent architecture, Min-Trick, and fallback mechanism. We address each major comment below and commit to revisions that strengthen the empirical support without altering the core claims.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: aggregate speedups are reported for MOT and LPT, yet the manuscript provides no statistics on fallback invocation frequency, post-Min-Trick duality gaps, or the correlation between per-instance prediction error and observed runtime reduction. Without these, it is impossible to determine whether RowDualNet consistently reduces LAPJV search effort on out-of-distribution data or whether gains are driven by a minority of instances.

    Authors: We agree that these statistics would improve transparency and allow better assessment of consistency. In the revised manuscript we will add: (i) fallback invocation rates on MOT and LPT, (ii) average and worst-case duality gaps after the Min-Trick adjustment, and (iii) a per-instance correlation analysis (scatter plot and coefficient) between prediction error and runtime reduction. This will clarify whether speedups are broadly distributed or driven by outliers on the real-world tasks. revision: yes

  2. Referee: [Method] Method and § on RowDualNet: because the network predicts duals independently per row, the paper should quantify how well the resulting (Min-Trick-adjusted) duals approximate the optimal dual solution on real-world instances; an ablation relating prediction error to search-tree pruning would directly test the central assumption that learned advice reliably accelerates the solver before fallback.

    Authors: We acknowledge that an explicit quantification of approximation quality would directly support the central hypothesis. While the current manuscript prioritizes end-to-end runtime and optimality guarantees, the revision will include an ablation on the MOT and LPT instances that reports the duality gap between Min-Trick-adjusted predictions and true optimal duals, together with a correlation between prediction error and LAPJV search-tree metrics (nodes explored or pruned). This ablation will be placed in the experimental section or appendix. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper trains RowDualNet on synthetic data to predict dual variables, then empirically measures runtime reductions on held-out synthetic and real-world instances while using a classical fallback (LAPJV) to enforce optimality. Speedups are reported as measured outcomes on unseen distributions, not derived by construction from the training objective. The Min-Trick is presented as a direct constructive adjustment from LP duality theory, independent of the neural predictions. No self-definitional equations, fitted parameters renamed as predictions, or load-bearing self-citations that collapse the central claim were identified. The framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the effectiveness of learned dual predictions and the correctness of the duality-based feasibility mechanism. No numerical free parameters are stated in the abstract.

axioms (1)
  • standard math The linear assignment problem admits a linear programming formulation with strong duality
    Invoked to justify the Min-Trick for constructive feasibility of predicted duals.
invented entities (2)
  • RowDualNet no independent evidence
    purpose: Lightweight row-independent neural architecture for predicting dual variables
    Introduced to avoid the O(N^2) memory cost of graph-based models while enabling large-scale warm-starts.
  • Min-Trick no independent evidence
    purpose: Constructive mechanism using LP duality to ensure feasibility of predicted duals without iterative projection
    New or repurposed device to eliminate costly projection steps.

pith-pipeline@v0.9.0 · 5591 in / 1405 out tokens · 90132 ms · 2026-05-12T03:17:12.060082+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages

  1. [1]

    Michael Mitzenmacher and Sergei Vassilvitskii , title =. Commun

    work page

  2. [2]

    Learning Predictions for Algorithms with Predictions , booktitle =

    Misha Khodak and Maria. Learning Predictions for Algorithms with Predictions , booktitle =

    work page

  3. [3]

    Binary Search with Distributional Predictions , booktitle =

    Michael Dinitz and Sungjin Im and Thomas Lavastida and Benjamin Moseley and Aidin Niaparast and Sergei Vassilvitskii , editor =. Binary Search with Distributional Predictions , booktitle =

    work page

  4. [4]

    Algorithms with Prediction Portfolios , booktitle =

    Michael Dinitz and Sungjin Im and Thomas Lavastida and Benjamin Moseley and Sergei Vassilvitskii , editor =. Algorithms with Prediction Portfolios , booktitle =

    work page

  5. [5]

    SeedGNN: Graph Neural Network for Supervised Seeded Graph Matching , booktitle =

    Liren Yu and Jiaming Xu and Xiaojun Lin , editor =. SeedGNN: Graph Neural Network for Supervised Seeded Graph Matching , booktitle =

    work page

  6. [6]

    Proceedings of the 40th International Conference on Machine Learning (

    Shinsaku Sakaue and Taihei Oki , title =. Proceedings of the 40th International Conference on Machine Learning (. 2023 , volume =

    work page 2023

  7. [7]

    Approximation algorithms for combinatorial optimisation with predictions , booktitle =

    Antonios Antoniadis and Marek Eli. Approximation algorithms for combinatorial optimisation with predictions , booktitle =

    work page

  8. [8]

    2025 , howpublished =

    work page 2025

  9. [9]

    2021 , howpublished =

    Tomas Kazmar , title =. 2021 , howpublished =

    work page 2021

  10. [10]

    Journal of mathematics and physics , IGNOREvolume=

    The distribution of a product from several sources to numerous localities , author=. Journal of mathematics and physics , IGNOREvolume=. 1941 , publisher=

    work page 1941

  11. [11]

    International Journal of Neural Systems , IGNOREvolume=

    A graph-based neural approach to linear sum assignment problems , author=. International Journal of Neural Systems , IGNOREvolume=. 2024 , publisher=

    work page 2024

  12. [12]

    International Conference on Data Mining (ICDM) , IGNOREpages=

    Computing optimal assignments in linear time for approximate graph matching , author=. International Conference on Data Mining (ICDM) , IGNOREpages=. 2019 , IGNOREorganization=

    work page 2019

  13. [13]

    Artificial intelligence and statistics (AISTATS) , IGNOREpages=

    Exploring the power of graph neural networks in solving linear optimization problems , author=. Artificial intelligence and statistics (AISTATS) , IGNOREpages=. 2024 , IGNOREorganization=

    work page 2024

  14. [14]

    Neural Information Processing Systems (

    Faster matchings via learned duals , author=. Neural Information Processing Systems (

    work page

  15. [15]

    Pattern Recognition , IGNOREvolume=

    GLAN: A graph-based linear assignment network , author=. Pattern Recognition , IGNOREvolume=. 2024 , publisher=

    work page 2024

  16. [16]

    IEEE Robotics and Automation Letters , IGNOREvolume=

    Graph neural network for decentralized multi-robot goal assignment , author=. IEEE Robotics and Automation Letters , IGNOREvolume=. 2024 , publisher=

    work page 2024

  17. [17]

    Symmetry , IGNOREvolume=

    HybridGNN: A Self-Supervised graph neural network for efficient maximum matching in bipartite graphs , author=. Symmetry , IGNOREvolume=. 2024 , publisher=

    work page 2024

  18. [18]

    Autonomous Agents and Multiagent Systems (AAMAS) , IGNOREpages=

    MAGNET: A Multi-Agent Graph Neural Network for Efficient Bipartite Task Assignment , author=. Autonomous Agents and Multiagent Systems (AAMAS) , IGNOREpages=

    work page

  19. [19]

    Results in Control and Optimization , IGNOREvolume=

    New auction algorithms for the assignment problem and extensions , author=. Results in Control and Optimization , IGNOREvolume=. 2024 , publisher=

    work page 2024

  20. [20]

    2024 , school=

    Solving Maximum Weighted Matching Problem Using Graph Neural Networks , author=. 2024 , school=

    work page 2024

  21. [21]

    International Conference on Learning Representations (

    Ziang Chen and Jialin Liu and Xinshang Wang and Wotao Yin , title =. International Conference on Learning Representations (

    work page

  22. [22]

    International Conference on Machine Learning (

    Faster fundamental graph algorithms via learned predictions , author=. International Conference on Machine Learning (. 2022 , IGNOREpublisher=

    work page 2022

  23. [23]

    Neural Information Processing Systems (

    Discrete-convex-analysis-based framework for warm-starting algorithms with predictions , author=. Neural Information Processing Systems (

    work page

  24. [24]

    Kuhn, Harold W , journal=. The. 1955 , publisher=

    work page 1955

  25. [25]

    Journal of the society for industrial and applied mathematics , IGNOREvolume=

    Algorithms for the assignment and transportation problems , author=. Journal of the society for industrial and applied mathematics , IGNOREvolume=. 1957 , publisher=

    work page 1957

  26. [26]

    Computing , IGNOREvolume=

    A shortest augmenting path algorithm for dense and sparse linear assignment problems , author=. Computing , IGNOREvolume=. 1987 , publisher=

    work page 1987

  27. [27]

    Improving the

    Jonker, Roy and Volgenant, Anton , journal=. Improving the. 1986 , publisher=

    work page 1986

  28. [28]

    SciPy 1.0: fundamental algorithms for scientific computing in

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and others , journal=. SciPy 1.0: fundamental algorithms for scientific computing in. 2020 , publisher=

    work page 2020

  29. [29]

    European Conference on Computer Vision (ECCV) , IGNOREpages=

    Large scale real-world multi-person tracking , author=. European Conference on Computer Vision (ECCV) , IGNOREpages=. 2022 , publisherpublisher=

    work page 2022

  30. [30]

    A graph-based neural approach to linear sum assignment problems

    Aironi, C., Cornell, S., and Squartini, S. A graph-based neural approach to linear sum assignment problems. International Journal of Neural Systems, 2024

    work page 2024

  31. [31]

    Approximation algorithms for combinatorial optimisation with predictions

    Antoniadis, A., Eli \' a s, M., Polak, A., and Venzin, M. Approximation algorithms for combinatorial optimisation with predictions. In International Conference on Learning Representations ( ICLR ) , 2025

    work page 2025

  32. [32]

    New auction algorithms for the assignment problem and extensions

    Bertsekas, D. New auction algorithms for the assignment problem and extensions. Results in Control and Optimization, 2024

    work page 2024

  33. [33]

    Faster fundamental graph algorithms via learned predictions

    Chen, J., Silwal, S., Vakilian, A., and Zhang, F. Faster fundamental graph algorithms via learned predictions. In International Conference on Machine Learning ( ICML ) , 2022

    work page 2022

  34. [34]

    On representing linear programs by graph neural networks

    Chen, Z., Liu, J., Wang, X., and Yin, W. On representing linear programs by graph neural networks. In International Conference on Learning Representations ( ICLR ) , 2023

    work page 2023

  35. [35]

    Faster matchings via learned duals

    Dinitz, M., Im, S., Lavastida, T., Moseley, B., and Vassilvitskii, S. Faster matchings via learned duals. Neural Information Processing Systems ( NeurIPS ) , 2021

    work page 2021

  36. [36]

    Algorithms with prediction portfolios

    Dinitz, M., Im, S., Lavastida, T., Moseley, B., and Vassilvitskii, S. Algorithms with prediction portfolios. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Neural Information Processing Systems ( NeurIPS ) , 2022

    work page 2022

  37. [37]

    Binary search with distributional predictions

    Dinitz, M., Im, S., Lavastida, T., Moseley, B., Niaparast, A., and Vassilvitskii, S. Binary search with distributional predictions. In Globersons, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J. M., and Zhang, C. (eds.), Neural Information Processing Systems ( NeurIPS ) , 2024

    work page 2024

  38. [38]

    Hitchcock, F. L. The distribution of a product from several sources to numerous localities. Journal of mathematics and physics, 1941

    work page 1941

  39. [39]

    and Volgenant, A

    Jonker, R. and Volgenant, A. Improving the H ungarian assignment algorithm. Operations research letters, 1986

    work page 1986

  40. [40]

    and Volgenant, A

    Jonker, R. and Volgenant, A. A shortest augmenting path algorithm for dense and sparse linear assignment problems. Computing, 1987

    work page 1987

  41. [41]

    LAP : L inear A ssignment P roblem solver

    Kazmar, T. LAP : L inear A ssignment P roblem solver. https://github.com/gatagat/lap, 2021

    work page 2021

  42. [42]

    Learning predictions for algorithms with predictions

    Khodak, M., Balcan, M., Talwalkar, A., and Vassilvitskii, S. Learning predictions for algorithms with predictions. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Neural Information Processing Systems ( NeurIPS ) , 2022

    work page 2022

  43. [43]

    M., Giscard, P.-L., Bause, F., and Wilson, R

    Kriege, N. M., Giscard, P.-L., Bause, F., and Wilson, R. C. Computing optimal assignments in linear time for approximate graph matching. In International Conference on Data Mining (ICDM), 2019

    work page 2019

  44. [44]

    Kuhn, H. W. The H ungarian method for the assignment problem. Naval research logistics quarterly, 1955

    work page 1955

  45. [45]

    Glan: A graph-based linear assignment network

    Liu, H., Wang, T., Lang, C., Feng, S., Jin, Y., and Li, Y. Glan: A graph-based linear assignment network. Pattern Recognition, 2024

    work page 2024

  46. [46]

    Magnet: A multi-agent graph neural network for efficient bipartite task assignment

    Loveland, D., Usevitch, J., Serlin, Z., Koutra, D., and Caceres, R. Magnet: A multi-agent graph neural network for efficient bipartite task assignment. In Autonomous Agents and Multiagent Systems (AAMAS), 2025

    work page 2025

  47. [47]

    and Vassilvitskii, S

    Mitzenmacher, M. and Vassilvitskii, S. Algorithms with predictions. Commun. ACM , 2022

    work page 2022

  48. [48]

    Algorithms for the assignment and transportation problems

    Munkres, J. Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics, 1957

    work page 1957

  49. [49]

    Metropolises of major cities around the World

    OpenStreetMap . Metropolises of major cities around the World . https://www.openstreetmap.org, 2025

    work page 2025

  50. [50]

    Hybridgnn: A self-supervised graph neural network for efficient maximum matching in bipartite graphs

    Pan, C.-H., Qu, Y., Yao, Y., and Wang, M.-J.-S. Hybridgnn: A self-supervised graph neural network for efficient maximum matching in bipartite graphs. Symmetry, 2024

    work page 2024

  51. [51]

    Exploring the power of graph neural networks in solving linear optimization problems

    Qian, C., Ch \'e telat, D., and Morris, C. Exploring the power of graph neural networks in solving linear optimization problems. In Artificial intelligence and statistics (AISTATS), 2024

    work page 2024

  52. [52]

    and Oki, T

    Sakaue, S. and Oki, T. Discrete-convex-analysis-based framework for warm-starting algorithms with predictions. Neural Information Processing Systems ( NeurIPS ) , 2022

    work page 2022

  53. [53]

    and Oki, T

    Sakaue, S. and Oki, T. Rethinking warm-starts with predictions: Learning predictions close to sets of optimal solutions for faster L- / L ^ -convex function minimization. In Proceedings of the 40th International Conference on Machine Learning ( ICML ) , volume 202 of PMLR, pp.\ 29760--29776, 2023

    work page 2023

  54. [54]

    Large scale real-world multi-person tracking

    Shuai, B., Bergamo, A., Buechler, U., Berneshawi, A., Boden, A., and Tighe, J. Large scale real-world multi-person tracking. In European Conference on Computer Vision (ECCV), 2022

    work page 2022

  55. [55]

    E., Haberland, M., Reddy, T., Cournapeau, D., Burovski, E., Peterson, P., Weckesser, W., Bright, J., et al

    Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau, D., Burovski, E., Peterson, P., Weckesser, W., Bright, J., et al. Scipy 1.0: fundamental algorithms for scientific computing in P ython. Nature methods, 2020

    work page 2020

  56. [56]

    Seedgnn: Graph neural network for supervised seeded graph matching

    Yu, L., Xu, J., and Lin, X. Seedgnn: Graph neural network for supervised seeded graph matching. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J. (eds.), International Conference on Machine Learning ( ICML ) , 2023

    work page 2023