arxiv: 2605.06588 · v1 · submitted 2026-05-07 · 💻 cs.LG · cs.AI

Recognition: unknown

Towards Metric-Faithful Neural Graph Matching

Jyotirmaya Shivottam , Subhankar Mishra

Authors on Pith no claims yet

Pith reviewed 2026-05-08 12:12 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords graph edit distanceneural graph matchingbi-Lipschitz encodersgraph neural networksGED estimationalignment costsranking stabilityFSW-GNN

0 comments

The pith

Bi-Lipschitz encoders produce controlled approximations to graph edit distance and more stable rankings in neural matchers.

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

Neural graph matching methods encode graphs with GNNs before either regressing a similarity score or optimizing a node alignment. This paper shows that when those encoders satisfy a bi-Lipschitz condition, the resulting GED estimates remain bounded by constants that depend on the encoder's Lipschitz constants, and ranking stability improves. The control holds on fixed collections where the doubly-stochastic metric stays comparable to true GED. For alignment-based estimators the same geometric property carries through to the node-level costs that are optimized. Replacing standard encoders with the bi-Lipschitz FSW-GNN in existing architectures yields measurable gains in prediction and ranking metrics across benchmarks.

Core claim

On fixed graph collections where the doubly-stochastic metric d_DS is comparable to GED, graph-level bi-Lipschitz encoders yield controlled GED surrogates and improved ranking stability; for matching-based estimators, node-level bi-Lipschitz geometry propagates to encoder-induced alignment costs and the resulting optimized alignment objective. The authors instantiate this perspective using FSW-GNN, a bi-Lipschitz WL-equivalent encoder, as a drop-in replacement and observe consistent improvements.

What carries the argument

bi-Lipschitz encoders whose embedding distances relate to input graph distances by multiplicative constants, thereby bounding approximation error and propagating geometry to both regression heads and alignment objectives.

If this is right

Graph-level regression heads produce GED surrogates whose error is bounded by the encoder's bi-Lipschitz constants.
Node alignment objectives inherit the same distance-control property, leading to more reliable optimized matchings.
Replacing encoders with FSW-GNN in existing neural GED architectures improves both prediction accuracy and ranking metrics.
Encoder geometry becomes a usable design principle for building new neural graph matchers.

Where Pith is reading between the lines

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

The same geometric constraint may improve neural estimators for other graph similarity measures once comparable reference metrics are identified.
Methods that explicitly optimize the Lipschitz constants of the encoder could extend the gains beyond the FSW-GNN instantiation.
The fixed-collection assumption indicates that performance on streaming or highly variable graph data would require additional verification or adaptive mechanisms.

Load-bearing premise

The results require fixed graph collections on which the doubly-stochastic metric remains comparable to GED and bi-Lipschitz encoders that can be realized without substantial loss of expressivity.

What would settle it

Measuring whether bi-Lipschitz encoders lose their advantage over standard encoders on graph collections where d_DS deviates substantially from GED would directly test the claimed propagation of geometric control.

Figures

Figures reproduced from arXiv: 2605.06588 by Jyotirmaya Shivottam, Subhankar Mishra.

**Figure 1.** Figure 1: Motivation for geometrically rigorous GNN encoders in view at source ↗

**Figure 2.** Figure 2: Faithfulness analysis of untrained GraSP view at source ↗

read the original abstract

Graph Edit Distance (GED) is a fundamental, albeit NP-hard, metric for structural graph similarity. Recent neural graph matching architectures approximate GED by first encoding graphs with a Graph Neural Network (GNN) and then applying either a graph-level regression head or a matching-based alignment module. Despite substantial architectural progress, the role of encoder geometry in neural GED estimation remains poorly understood. In this paper, we develop a theoretical framework that connects encoder geometry to GED estimation quality for two broad classes of neural GED estimators: graph similarity predictors and alignment-based methods. On fixed graph collections, where the doubly-stochastic metric $d_{\mathrm{DS}}$ is comparable to GED, we show that graph-level bi-Lipschitz encoders yield controlled GED surrogates and improved ranking stability; for matching-based estimators, node-level bi-Lipschitz geometry propagates to encoder-induced alignment costs and the resulting optimized alignment objective. We instantiate this perspective using FSW-GNN, a bi-Lipschitz WL-equivalent encoder, as a drop-in replacement in representative neural GED architectures. Across representative baselines and benchmark datasets, the resulting geometry-aware variants significantly improve GED prediction and ranking metrics. A faithfulness case study of untrained encoders, together with ablations and transfer experiments, supports the view that these gains arise from improved representation geometry, positioning encoder geometry as a useful design principle for neural graph matching.

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.

Desk Editor's Note private letter to a colleague

The paper connects bi-Lipschitz encoder geometry to improved neural GED estimation, but the key d_DS comparability assumption lacks explicit verification.

read the letter

The one thing to know is that this work ties bi-Lipschitz properties of graph encoders to more reliable neural estimates of graph edit distance, with a new model called FSW-GNN as the vehicle. They lay out a framework that explains why certain encoder geometries help with both regression-style GED predictors and alignment-based matchers. On fixed collections where d_DS tracks GED, the bi-Lipschitz condition gives bounds on the surrogate error and better ranking. For matching, it carries through to the alignment costs. FSW-GNN is presented as WL-equivalent yet bi-Lipschitz, and swapping it into baselines yields measurable lifts in accuracy and stability. The faithfulness study on untrained models is a nice touch to separate geometry from optimization effects. The main limitation is that everything hinges on d_DS being a good proxy for GED in the chosen datasets. The abstract and claims treat this as holding, but there is no clear evidence of verification or quantitative bounds in the provided details. If the two metrics diverge on structural features the benchmarks care about, the claimed improvements in surrogates and objectives lose their grounding. The empirical gains are real, but attributing them cleanly to the geometry requires that assumption to check out. This paper is aimed at the graph learning community focused on similarity and matching tasks, particularly those applying it to chemistry or biology graphs. A reader who cares about representation properties in GNNs rather than just bigger models will find it relevant. It has enough substance in the theory and experiments to merit a serious referee, though the conditional aspect means it is not a slam dunk. I would send it to peer review. The core idea is worth the community's attention, and referees can push on the comparability point and the proof details.

Referee Report

3 major / 2 minor

Summary. The manuscript develops a theoretical framework connecting encoder geometry (specifically bi-Lipschitz properties at graph and node levels) to the quality of neural Graph Edit Distance (GED) estimation in two classes of architectures: graph-level regression predictors and matching-based alignment methods. On fixed graph collections where the doubly-stochastic metric d_DS is comparable to GED, it claims that graph-level bi-Lipschitz encoders yield controlled GED surrogates with improved ranking stability, while node-level bi-Lipschitz geometry propagates to encoder-induced alignment costs and optimized objectives. The framework is instantiated via FSW-GNN (a bi-Lipschitz WL-equivalent encoder) as a drop-in replacement, with empirical gains reported across baselines and benchmarks, supported by a faithfulness case study on untrained encoders, ablations, and transfer experiments.

Significance. If the theoretical connections and the d_DS comparability hold, the work offers a principled design principle for neural graph matching by prioritizing metric-faithful geometry over pure expressivity, which could guide future encoder choices. The empirical improvements, the untrained-encoder case study, and the distinction between graph-level and node-level effects are strengths that add insight beyond standard benchmarks. The conditional framing, however, narrows the immediate applicability until the key assumption is verified.

major comments (3)

[Theoretical framework and abstract claims] The central theoretical claims (graph-level bi-Lipschitz encoders yielding controlled GED surrogates; node-level geometry propagating to alignment costs) are explicitly conditioned on d_DS being comparable to GED on the fixed collections used. No explicit verification, quantitative bounds, or empirical checks confirming this comparability on the specific benchmark graphs appear in the theoretical framework or experimental sections; without this, the propagation results do not follow even if FSW-GNN is bi-Lipschitz and WL-equivalent.
[Theoretical results and derivations] The abstract and strongest claim reference 'theoretical results' on bi-Lipschitz propagation, yet the manuscript provides no full derivations or proofs establishing how graph-level or node-level bi-Lipschitz geometry maps to GED surrogates and optimized alignment objectives. This absence prevents verification of the load-bearing steps in the framework.
[Experiments and faithfulness case study] The empirical gains and attribution to improved representation geometry (via the faithfulness case study and ablations) rest on the unverified d_DS-GED comparability; if the proxy fails on the benchmarks, alternative explanations for the ranking and prediction improvements cannot be ruled out.

minor comments (2)

[Introduction] Notation for d_DS and its relation to GED could be introduced with a brief reminder equation in the introduction for readers unfamiliar with the doubly-stochastic metric.
[Instantiation of FSW-GNN] The description of FSW-GNN as 'drop-in replacement' would benefit from a short table comparing its properties (bi-Lipschitz constant, WL equivalence) to the original encoders it replaces.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments correctly identify that our theoretical claims are conditional and that the manuscript would benefit from explicit verification and fuller derivations. We address each point below and will revise the manuscript accordingly to strengthen the presentation.

read point-by-point responses

Referee: [Theoretical framework and abstract claims] The central theoretical claims (graph-level bi-Lipschitz encoders yielding controlled GED surrogates; node-level geometry propagating to alignment costs) are explicitly conditioned on d_DS being comparable to GED on the fixed collections used. No explicit verification, quantitative bounds, or empirical checks confirming this comparability on the specific benchmark graphs appear in the theoretical framework or experimental sections; without this, the propagation results do not follow even if FSW-GNN is bi-Lipschitz and WL-equivalent.

Authors: We agree that the claims are conditional on d_DS-GED comparability and that explicit checks on the specific benchmarks would improve rigor. The benchmarks were selected from collections where prior work has established d_DS as a reliable GED surrogate, but we did not include quantitative verification in the original manuscript. In revision we will add an appendix containing empirical comparisons (Pearson/Spearman correlations, maximum deviations, and bound checks) between d_DS and GED on all datasets used in the experiments. revision: yes
Referee: [Theoretical results and derivations] The abstract and strongest claim reference 'theoretical results' on bi-Lipschitz propagation, yet the manuscript provides no full derivations or proofs establishing how graph-level or node-level bi-Lipschitz geometry maps to GED surrogates and optimized alignment objectives. This absence prevents verification of the load-bearing steps in the framework.

Authors: The framework connects bi-Lipschitz conditions to the resulting surrogate bounds via the metric definitions and Lipschitz constants, but we acknowledge that the main text presents these connections at a high level without complete step-by-step derivations. We will add a dedicated appendix with full proofs for both the graph-level case (controlled GED surrogates and ranking stability) and the node-level case (propagation to alignment costs and optimized objectives). revision: yes
Referee: [Experiments and faithfulness case study] The empirical gains and attribution to improved representation geometry (via the faithfulness case study and ablations) rest on the unverified d_DS-GED comparability; if the proxy fails on the benchmarks, alternative explanations for the ranking and prediction improvements cannot be ruled out.

Authors: We will incorporate the d_DS-GED comparability verification described in our response to the first comment. This addition will provide direct support for attributing the observed improvements to the metric-faithful encoder geometry, reinforcing the existing faithfulness case study, ablations, and transfer experiments. revision: yes

Circularity Check

0 steps flagged

No significant circularity; theoretical claims are conditional and self-contained.

full rationale

The paper's core derivation is a conditional theoretical analysis: on fixed collections where d_DS is comparable to GED, graph-level bi-Lipschitz encoders yield controlled GED surrogates, and node-level bi-Lipschitz geometry propagates to alignment costs. This is derived from the stated bi-Lipschitz and WL-equivalence properties of the encoder (instantiated as FSW-GNN) without reducing any result to a fitted parameter, self-definition, or self-citation chain. The comparability assumption is explicitly flagged as a precondition rather than derived internally. Empirical results on benchmarks are presented as validation of the geometry-aware variants, not as the load-bearing step that forces the theoretical claims. No steps match the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The paper rests on the domain assumption that d_DS approximates GED on fixed collections and introduces the new entity FSW-GNN; no free parameters are mentioned in the abstract.

axioms (1)

domain assumption On fixed graph collections the doubly-stochastic metric d_DS is comparable to GED.
Explicitly stated as the setting in which the bi-Lipschitz results hold.

invented entities (1)

FSW-GNN no independent evidence
purpose: A bi-Lipschitz WL-equivalent encoder used as drop-in replacement in neural GED architectures.
Introduced in the paper as the concrete instantiation of the proposed geometry-aware approach.

pith-pipeline@v0.9.0 · 5537 in / 1464 out tokens · 98157 ms · 2026-05-08T12:12:01.632363+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 15 canonical work pages

[1]

FSW-GNN: A Bi-Lipschitz WL-Equivalent Graph Neural Network

Yonatan Sverdlov, Yair Davidson, Nadav Dym, and Tal Amir. FSW-GNN: A Bi-Lipschitz WL-Equivalent Graph Neural Network. InThe Fourth Learning on Graphs Conference, October 2025

2025
[2]

Runzhong Wang, Rui-Xi Wang, Mrunali Manjrekar, and Connor W. Coley. Neural Graph Matching Improves Retrieval Augmented Generation in Molecular Machine Learning. In Forty-Second International Conference on Machine Learning, June 2025

2025
[3]

Graph Edit Distance Estimation: A New Heuristic and A Holistic Evaluation of Learning-based Methods.Proc

Mouyi Xu and Lijun Chang. Graph Edit Distance Estimation: A New Heuristic and A Holistic Evaluation of Learning-based Methods.Proc. ACM Manag. Data, 3(3):167:1–167:24, June
[4]

doi: 10.1145/3725304

work page doi:10.1145/3725304
[5]

Graph Theory-Based Deep Graph Similarity Learning: A Unified Survey of Pipeline, Techniques, and Challenges.Transactions on Machine Learning Research, December 2024

Zhouyang Liu, Ning Liu, Yixin Chen, Ziqing Wen, Jiezhong He, and Dongsheng Li. Graph Theory-Based Deep Graph Similarity Learning: A Unified Survey of Pipeline, Techniques, and Challenges.Transactions on Machine Learning Research, December 2024. ISSN 2835-8856

2024
[6]

Approximate graph edit distance computation by means of bipartite graph matching.Image and Vision Computing, 27(7):950–959, 2009

Kaspar Riesen and Horst Bunke. Approximate graph edit distance computation by means of bipartite graph matching.Image and Vision Computing, 27(7):950–959, 2009. ISSN 0262-8856. doi: 10.1016/j.imavis.2008.04.004

work page doi:10.1016/j.imavis.2008.04.004 2009
[7]

Grasp: simple yet effective graph similarity predictions

Haoran Zheng, Jieming Shi, and Renchi Yang. Grasp: simple yet effective graph similarity predictions. InProceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence, AAAI’25/IAAI’25/EA...

work page doi:10.1609/aaai.v39i21.34450 2025
[8]

Fishkind, Marcelo Fiori, Joshua T

Vince Lyzinski, Donniell E. Fishkind, Marcelo Fiori, Joshua T. V ogelstein, Carey E. Priebe, and Guillermo Sapiro. Graph matching: Relax at your own risk.IEEE Trans. Pattern Anal. Mach. Intell., 38(1):60–73, January 2016. ISSN 0162-8828. doi: 10.1109/TPAMI.2015.2424894. URLhttps://doi.org/10.1109/TPAMI.2015.2424894

work page doi:10.1109/tpami.2015.2424894 2016
[9]

Ds: Tighter lifting-free convex relaxations for quadratic matching problems

Florian Bernard, Christian Theobalt, and Michael Moeller. Ds: Tighter lifting-free convex relaxations for quadratic matching problems. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310–4319, 2018. doi: 10.1109/CVPR.2018.00453

work page doi:10.1109/cvpr.2018.00453 2018
[10]

Graph Matching Networks for Learning the Similarity of Graph Structured Objects

Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph Matching Networks for Learning the Similarity of Graph Structured Objects. InProceedings of the 36th International Conference on Machine Learning, pages 3835–3845. PMLR, May 2019

2019
[11]

Simgnn: A neural network approach to fast graph similarity computation

Yunsheng Bai, Hao Ding, Song Bian, Ting Chen, Yizhou Sun, and Wei Wang. Simgnn: A neural network approach to fast graph similarity computation. InProceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM ’19, page 384–392, New York, NY , USA, 2019. Association for Computing Machinery. ISBN 9781450359405. doi: 10.1145/3289...

work page doi:10.1145/3289600.3290967 2019
[12]

Computing Graph Edit Distance via Neural Graph Matching.Proceedings of the VLDB Endowment, 16(8):1817–1829, April 2023

Chengzhi Piao, Tingyang Xu, Xiangguo Sun, Yu Rong, Kangfei Zhao, and Hong Cheng. Computing Graph Edit Distance via Neural Graph Matching.Proceedings of the VLDB Endowment, 16(8):1817–1829, April 2023. ISSN 2150-8097. doi: 10.14778/3594512.3594514

work page doi:10.14778/3594512.3594514 2023
[13]

Graph Edit Distance with General Costs Using Neural Set Divergence

Eeshaan Jain, Indradyumna Roy, Saswat Meher, Soumen Chakrabarti, and Abir De. Graph Edit Distance with General Costs Using Neural Set Divergence. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, November 2024. 10

2024
[14]

Fused Gromov-Wasserstein Alignment for Graph Edit Distance Computation and Beyond.Proc

Jianheng Tang, Xi Zhao, Lemin Kong, Xiaofang Zhou, and Jia Li. Fused Gromov-Wasserstein Alignment for Graph Edit Distance Computation and Beyond.Proc. VLDB Endow., 18(10): 3641–3654, June 2025. ISSN 2150-8097. doi: 10.14778/3748191.3748221

work page doi:10.14778/3748191.3748221 2025
[15]

How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019. URL https: //openreview.net/forum?id=ryGs6iA5Km

2019
[16]

On the Hölder Stability of Multiset and Graph Neural Networks

Yair Davidson and Nadav Dym. On the Hölder Stability of Multiset and Graph Neural Networks. InThe Thirteenth International Conference on Learning Representations, October 2024

2024
[17]

Posi- tion: Graph Matching Systems Deserve Better Benchmarks

Indradyumna Roy, Saswat Meher, Eeshaan Jain, Soumen Chakrabarti, and Abir De. Posi- tion: Graph Matching Systems Deserve Better Benchmarks. InForty-Second International Conference on Machine Learning Position Paper Track, June 2025

2025
[18]

GREED: A Neural Framework for Learning Graph Distance Functions

Rishabh Ranjan, Siddharth Grover, Sourav Medya, Venkatesan Chakaravarthy, Yogish Sabhar- wal, and Sayan Ranu. GREED: A Neural Framework for Learning Graph Distance Functions. InAdvances in Neural Information Processing Systems, October 2022

2022
[19]

Iverson, Dustin G

Jameson Cahill, Joseph W. Iverson, Dustin G. Mixon, and Nathan Willey. Approximation theorems in bilipschitz invariant theory, March 2026

2026
[20]

Gelato: Graph Edit Distance via Autoregressive Neural Combinatorial Optimization

Paolo Pellizzoni, Till Hendrik Schulz, and Karsten Borgwardt. Gelato: Graph Edit Distance via Autoregressive Neural Combinatorial Optimization. InThe Fourteenth International Conference on Learning Representations, October 2025

2025
[21]

Yunsheng Bai, Hao Ding, Ken Gu, Yizhou Sun, and Wei Wang. Learning-Based Efficient Graph Similarity Computation via Multi-Scale Convolutional Set Matching.Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):3219–3226, April 2020. ISSN 2374-3468. doi: 10.1609/aaai.v34i04.5720

work page doi:10.1609/aaai.v34i04.5720 2020
[22]

Slow learning and fast inference: Efficient graph similarity computation via knowledge distillation

Can Qin, Handong Zhao, Lichen Wang, Huan Wang, Yulun Zhang, and Yun Fu. Slow learning and fast inference: Efficient graph similarity computation via knowledge distillation. In M. Ran- zato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 14110–14121. Curran Associ...

2021
[23]

Efficient graph similarity computation with alignment regularization

Wei Zhuo and Guang Tan. Efficient graph similarity computation with alignment regularization. InProceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY , USA, 2022. Curran Associates Inc. ISBN 9781713871088

2022
[24]

H2mn: Graph similarity learning with hierarchical hypergraph matching networks

Zhen Zhang, Jiajun Bu, Martin Ester, Zhao Li, Chengwei Yao, Zhi Yu, and Can Wang. H2mn: Graph similarity learning with hierarchical hypergraph matching networks. InPro- ceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Min- ing, KDD ’21, page 2274–2284, New York, NY , USA, 2021. Association for Comput- ing Machinery. ISBN 9781450383...

work page doi:10.1145/3447548.3467328 2021
[25]

Interpretable neural subgraph matching for graph retrieval.Proceedings of the AAAI Conference on Artificial Intelligence, 36(7):8115–8123, Jun

Indradyumna Roy, Venkata Sai Baba Reddy Velugoti, Soumen Chakrabarti, and Abir De. Interpretable neural subgraph matching for graph retrieval.Proceedings of the AAAI Conference on Artificial Intelligence, 36(7):8115–8123, Jun. 2022. doi: 10.1609/aaai.v36i7.20784. URL https://ojs.aaai.org/index.php/AAAI/article/view/20784

work page doi:10.1609/aaai.v36i7.20784 2022
[26]

GEDAN: Learning the edit costs for graph edit distance, 2026

Francesco Leonardi, Markus Orsi, Jean-Louis Reymond, and Kaspar Riesen. GEDAN: Learning the edit costs for graph edit distance, 2026. URL https://openreview.net/forum?id= Y5icGPy7yc

2026
[27]

Tagsim: type-aware graph similarity learning and computation

Jiyang Bai and Peixiang Zhao. Tagsim: type-aware graph similarity learning and computation. Proc. VLDB Endow., 15(2):335–347, October 2021. ISSN 2150-8097. doi: 10.14778/3489496. 3489513. URLhttps://doi.org/10.14778/3489496.3489513. 11

work page doi:10.14778/3489496 2021
[28]

GRAIL: Graph edit distance and node alignment using LLM-generated code

Samidha Verma, Arushi Goyal, Ananya Mathur, Ankit Anand, and Sayan Ranu. GRAIL: Graph edit distance and node alignment using LLM-generated code. InForty-second Inter- national Conference on Machine Learning, 2025. URLhttps://openreview.net/forum? id=NzoZXju2bL

2025
[29]

Ds++: a flexible, scalable and provably tight relaxation for matching problems.ACM Trans

Nadav Dym, Haggai Maron, and Yaron Lipman. Ds++: a flexible, scalable and provably tight relaxation for matching problems.ACM Trans. Graph., 36(6), November 2017. ISSN 0730-0301. doi: 10.1145/3130800.3130826. URLhttps://doi.org/10.1145/3130800.3130826

work page doi:10.1145/3130800.3130826 2017
[30]

Tree movers distance: Bridging graph metrics and stability of graph neural networks

Ching-Yao Chuang and Stefanie Jegelka. Tree movers distance: Bridging graph metrics and stability of graph neural networks. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 2944–2957. Curran Associates, Inc., 2022

2022
[31]

Quan- titative Bounds for Sorting-Based Permutation-Invariant Embeddings.IEEE Transactions on Information Theory, pages 1–1, 2026

Nadav Dym, Matthias Wellershoff, Efstratios Tsoukanis, Daniel Levy, and Radu Balan. Quan- titative Bounds for Sorting-Based Permutation-Invariant Embeddings.IEEE Transactions on Information Theory, pages 1–1, 2026. ISSN 1557-9654. doi: 10.1109/TIT.2026.3679460

work page doi:10.1109/tit.2026.3679460 2026
[32]

3-WL GNNs for Metric Learning on Graphs

Aldo Moscatelli, Maxime Bérar, Pierre Héroux, Florian Yger, and Sébastien Adam. 3-WL GNNs for Metric Learning on Graphs. InESANN 2025, Brugges, Belgium, April 2025. URL https://hal.science/hal-05327901

2025
[34]

URLhttp://arxiv.org/abs/1711.07553

work page arXiv
[35]

Reasoning with neu- ral tensor networks for knowledge base completion

Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. Reasoning with neu- ral tensor networks for knowledge base completion. In C.J. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger, editors,Advances in Neural Information Processing Sys- tems, volume 26. Curran Associates, Inc., 2013. URL https://proceedings.neurips.cc/ paper_fi...

2013
[36]

Gromov-Wasserstein Learning for Graph Matching and Node Embedding

Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-Wasserstein Learning for Graph Matching and Node Embedding. InProceedings of the 36th International Conference on Machine Learning, pages 6932–6941. PMLR, May 2019. A Broader Impact This work is primarily foundational. It studies how encoder geometry affects the quality of neural graph ...

2019