Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Target Circuit Matching in Large-Scale Netlists using GNN-Based Region Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A GNN trained to rank K-hop neighborhoods by probability of containing a target graph lets exact VF2 matching find every target occurrence in large netlists while cutting runtime by up to 96.47% versus running VF2 on the whole graph.

desk verdict A plausible applied-GNN paper with one genuinely nice embedding trick, but the headline speedup is not yet pinned down because the ranked-search stopping rule is unspecified and ranking quality on unseen circuits is never measured. read the letter →

arxiv 2507.19518 v1 pith:43R553XB submitted 2025-07-18 cs.LG cs.AI

classification cs.LGcs.AI
keywords subgraphmatchinggraphneuralnetworkstargetregionpredictionK-hopembeddingVF2negativesamplingcircuitverificationnetlistrepresentation
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

The paper proposes a graph neural network approach for exact subgraph matching in large circuits. Instead of applying a rule-based matcher over the entire netlist or optimizing a node-to-node matching matrix, it trains a GNN to score each K-hop neighborhood of the circuit graph by the probability that it fully contains a given target graph, then feeds the highest-scoring neighborhoods to the exact VF2 algorithm. The authors claim this finds the same number of matches as plain VF2 on three large unseen circuits while reducing runtime by up to 96.47%, and that the region-ranking component beats prior graph-matching networks in accuracy and AUROC. The practical interest is that exact subgraph matching becomes affordable for large-scale EDA verification tasks such as fault localization and netlist hierarchy generation.

What carries the argument

The central object is the K-hop subgraph embedding $\mathbf{h}_{K\text{-hop}} = \operatorname{Concat}(\mathbf{h}^{(0)}_{K\text{-hop}}, \ldots, \mathbf{h}^{(L)}_{K\text{-hop}})$, where $\mathbf{h}^{(l)}_{K\text{-hop}}$ is the sum of the $l$-th layer node embeddings over the $(K-l)$-hop neighborhood of the center node. Since the GNN is run once on the entire graph and each layer's embeddings at distance $K-l$ or less are pooled separately, the concatenation contains exactly the information inside the K-hop ball. $K$ is set to the radius of the target graph, the smallest hop count that can enclose it. An MLP maps this embedding together with the target graph embedding to a probability that the target is present, and VF2 verifies candidates in descending probability order. The four negative sample types (Partial, Mutation, Others, Random) are what force the classifier to separate full containment from near-misses.

What would settle it

Re-running the reported evaluation but replacing the GNN's predicted probabilities with random scores, or with a constant score, while keeping the VF2 verification step unchanged would settle the claim: if the speedup over plain VF2 persists under random ranking, the gain comes from VF2's own pruning on K-hop subgraphs rather than from the learned ranking; if random ranking is much slower, the learned ranking is the load-bearing component.

Watch

Extended reading notes

Core claim

The central claim is that expensive exact subgraph matching can be made scalable by framing it as a ranking problem: a GNN-based classifier predicts which K-hop subgraphs of a large circuit graph contain the target graph, and VF2, run only on those ranked regions in descending order of predicted probability, recovers all target occurrences exactly. The method represents MOSFET netlists as directed typed graphs with one node per cell, uses relational graph convolutions with basis decomposition applied once to the whole circuit, and derives every candidate subgraph embedding by concatenating per-hop pooled embeddings $\mathbf{h}^{(l)}_{K\text{-hop}} = \operatorname{Pool}(\{\mathbf{h}^{(l)}_i \mid \operatorname{dist}(v_c, v_i) \le K-l\})$ so that no information from outside the K-hop boundary leaks into the representation. Training mixes positive samples with four types of negatives, including hard negatives that partially contain the target or mutate PMOS/NMOS types, which the paper shows is necessary for the speedup. Empirically, on the three largest circuits used for evaluation, the method matches every occurrence found by VF2 and cuts matching time by up to 96.47% for all-target matching, with single-target time reduced by more than 98% relative to the nearest baselines in most cases.

Load-bearing premise

The speedup assumes the learned ranking stays accurate on circuits and target graphs the model has not seen; if the ranking were no better than random, VF2 would be run on nearly every K-hop neighborhood and the whole procedure could be slower than running VF2 once on the full circuit.

Editorial extensions

If this is right

  • On unseen large circuits, exact matching finds every target occurrence that plain VF2 finds, so the ranking does not trade away recall.
  • All-target matching time drops by up to 96.47% compared with whole-graph VF2, and single-target matching beats FastPFP and NeuroMatch by over 98% in most tested cases.
  • Computing one set of node embeddings for the whole circuit keeps GNN cost independent of the number of candidate K-hop subgraphs; the ablation shows this is what removes the per-candidate bottleneck.
  • Hard negative samples (Partial and Mutation) are responsible for much of the speedup; ablations with only random negatives run slower.
  • The resulting pipeline is ready for fault-region identification and netlist hierarchy generation in large MOSFET netlists.

Reading between the lines

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

  • A test the paper does not run would be to measure region-prediction AUROC on the unseen target circuits used in Table II (col_sel1m, decoder2_4, deco1m); that would separate ranking quality on unseen targets from VF2-side pruning effects.
  • Because the ranking layer is independent of the exact matcher, pairing the same region predictor with a faster exact subgraph-isomorphism engine such as VF3 would likely compound the speedup.
  • The typed directed graph encoding and the whole-graph K-hop embedding extraction could transfer to other netlist learning tasks such as cell classification or layout-aware feature prediction, not only to matching.
  • For very small targets like inv, the reported speedup is small (2.81%), consistent with the paper's observation that small targets are harder to distinguish; raising K or adding structural fingerprints for small graphs is a natural extension.
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 / 6 minor

Summary. The paper proposes a GNN-based method for exact subgraph matching in large circuit netlists. It represents MOSFET netlists as directed heterogeneous graphs, trains a relational GNN with basis decomposition as a binary classifier on K-hop subgraphs (positive samples containing the target, plus four negative types), and at inference ranks all K-hop subgraphs centered at every node by predicted probability of containing the target, then runs VF2 on the ranked subgraphs. Experiments on the Lassen DRAMB dataset compare against VF2 for all-occurrence matching and against FastPFP and NeuroMatch for single-occurrence matching, and report target region prediction accuracy/AUROC against GNN and graph similarity baselines.

Significance. If the efficiency claim holds, the method could be a practical advance for EDA subcircuit identification because it scales to large netlists where plain VF2 is prohibitive and avoids node-to-node matching matrices. The K-hop subgraph embedding extraction trick (Section III-C2, Eqs. (6)-(7)) is a clean way to avoid per-subgraph GNN forward passes, and the ablation in Table VI supports that component. The paper provides source code. However, the central runtime comparison is currently confounded by the unspecified termination rule and the absence of a non-learned K-hop baseline, and the generalization evidence for ranking quality on unseen target circuits is indirect.

major comments (4)
  1. [§IV-A, Table II] The all-occurrence timing comparison in Table II does not state how the ranked search terminates. The proposed pipeline in §III-D sorts K-hop subgraphs by predicted probability and applies VF2 in that order, but the paper never specifies how the method knows that all target occurrences have been found. If the implementation processes every center node, the learned ordering has no effect on the all-occurrence runtime; if it stops after finding the same match count as the VF2 baseline, the comparison is circular because that count is unknown in real use. The reported up-to-96.47% runtime reduction is therefore not yet established for practical all-occurrence matching. The authors should specify the stopping criterion and report results under a protocol that does not use the ground-truth count (e.g., a probability threshold or a time budget), or compare with an exhaustive K-hop baseline that does not use ranking.
  2. [§IV-A, Table II] The comparison against plain VF2 on the whole graph conflates the benefit of the learned ranking with the benefit of restricting matching to K-hop subgraphs. Since each K-hop subgraph is smaller than the full graph, running VF2 on every K-hop subgraph without any learned ordering may already be faster than VF2 on the whole graph. A non-learned control—for instance, VF2 on all K-hop subgraphs in arbitrary or unranked order—is needed to attribute the speedup to the GNN-based region prediction. Without such a control, the claim that the method 'significantly outperforms existing methods in terms of time efficiency' is not fully supported.
  3. [§IV-B and Evaluation Protocol] Target region prediction is evaluated in Table V only for the eight training target circuits and for positive/random samples drawn from the training entire circuits. No AUROC or accuracy is reported for the unseen entire circuits used in Table II (col_sel9_512, deco9_512, ctrl256kbm) or for the unseen target graphs of Table II (col_sel1m, decoder2_4, deco1m). The only evidence that the ranking generalizes to these settings is the final runtime, which is confounded by the termination issue noted above. Please report region-prediction quality on the exact target/entire-circuit combinations of Table II, or explicitly acknowledge that ranking generalization on those circuits is unmeasured.
  4. [§IV-A, Table II] For the small abundant target inv, the runtime reduction over VF2 is only 2.81% (238.02s vs 231.33s), consistent with the authors' observation in §IV-B that smaller targets are harder to distinguish. Because small targets such as inv are common in circuit verification, this case substantially limits the practical significance of the efficiency claim. The paper should discuss this limitation and, if possible, characterize how the speedup depends on target size and occurrence count.
minor comments (6)
  1. [Table II] Table II is difficult to read because the entire-circuit labels are separated from their data rows by the column layout; please reformat so each block of rows is unambiguously associated with its entire circuit.
  2. [Tables II, IV, VI] Timing results in Tables II, IV, and VI are reported without standard deviations or number of runs; given that Table V reports mean±std over seeds, please provide similar repeated-run statistics for the runtime comparisons.
  3. [§IV, Evaluation Protocol] Several hyperparameters used in the experiments are not specified, including the embedding dimension, the number of basis matrices B in Eq. (4), the MLP architecture, and how K is computed in practice from the target radius.
  4. [References] Reference [23] appears to have the wrong title: 'A high-performance subcircuit recognition method based on the nonlinear graph optimization' is the title of Rubanov's paper [26]; please correct the citation.
  5. [§III-C2] The notation 'V^l_K(vc)' and 'h(l)K-hop' is introduced without defining how the target-graph embeddings in Eq. (8) are computed—whether the GNN is applied separately to the target graph or extracted from the entire-graph pass; please clarify.
  6. [§IV, Evaluation Protocol] In the Evaluation Protocol, the training set includes 18 entire circuits, but the list appears to contain more than 18 names; please verify the count and the set membership.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GNN ranking plus VF2 verification is a self-contained empirical pipeline; timing caveats are evaluation concerns, not circularity.

full rationale

No load-bearing circularity found. The method trains a GNN classifier on positive and negative K-hop subgraphs (Sec. III-C), then uses the predicted probabilities only to order VF2 verification calls (Sec. III-D). Prediction and verification are separate: matches are counted by VF2, not by the GNN, so the reported equality of match counts with VF2 is an empirical correctness property of exhaustive verification, not a fitted input. The target-region prediction evaluation (Table V) uses positive and random samples, and the main matching evaluation uses three entire circuits (ctrl256kbm, deco9_512, col_sel9_512) that are explicitly stated to be unseen during training, providing external evidence for the learned ranking. The principal caveats are evaluation-protocol ambiguities, not circularity: the paper does not state a stopping rule for the all-target matching timing runs, and Table V's negative samples are only random subgraphs, so ranking quality on the unseen target circuits in Table II is not directly measured. These caveats affect the strength of the efficiency claims but do not make any derivation reduce to its own inputs. There is no self-citation chain, no imported uniqueness theorem, and no equation-level construction that equates a prediction with a fitted parameter.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the trained GNN's ranking quality, the graph representation, and VF2's correctness. No new physical entities are introduced. The main unpurchased assumptions are that the representation preserves matching semantics and that the random-negative labeling is clean.

free parameters (7)
  • GNN weights and basis decomposition parameters = learned from training data
    All model parameters (W_r^(l), V_b^(l), a_rb^(l)) are fit to the training examples; the matching pipeline depends on them.
  • Hop radius K = Radius(GT) for each target
    Set to the radius of the target graph; defines the subgraph size and is central to the method.
  • Number of GNN layers L = 2
    Chosen by the authors for all experiments.
  • Number of basis matrices B = not reported
    The basis decomposition in Eq. 4 introduces B, but the paper never states its value.
  • Negative sample counts = 200 Pos, 100 Partial, 50 Mutation, 50 Others, 300 Random per target
    Hand-chosen ratios for training data construction.
  • Learning rate and epochs = 0.001, 1000
    Standard optimizer settings; not varied.
  • MLP head parameters = learned
    The final classifier in Eq. 10 adds its own fitted weights.
assumptions (5)
  • standard math Triangle inequality for graph distances ensures that pooling layer-l embeddings over (K-l)-hop neighborhoods excludes out-of-boundary information.
    Section III-C2 relies on dist(vc,j) <= dist(vc,i) + dist(i,j) <= K for i within K-l hops.
  • domain assumption The directed graph representation with one node per MOSFET and 28 edge types is sufficient to distinguish terminal roles for matching.
    Section III-A; if this representation loses information needed for isomorphism, matches could be missed.
  • standard math VF2 correctly finds all exact subgraph isomorphisms.
    Used as the verification step in Section III-D; correctness of match counts depends on VF2.
  • domain assumption The Lassen DRAMB netlists are representative of large-scale circuits for generalization.
    Evaluation is on three DRAM circuits from one dataset; generalization to other circuit families is untested.
  • domain assumption Random K-hop subgraphs not containing GT are correctly labeled negative.
    Section III-B samples Random negatives except positive samples; for small targets like inv appearing thousands of times, random subgraphs can contain GT and would be mislabeled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Target Circuit Matching in Large-Scale Netlists using GNN-Based Region Prediction." pith.science (2026). https://pith.science/paper/43R553XB

@misc{pith2026250719518,
  author       = {Pith},
  title        = {Pith review of: Target Circuit Matching in Large-Scale Netlists using GNN-Based Region Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/43R553XB}},
  note         = {Machine review of arXiv:2507.19518}
}
read the original abstract

Subgraph matching plays an important role in electronic design automation (EDA) and circuit verification. Traditional rule-based methods have limitations in generalizing to arbitrary target circuits. Furthermore, node-to-node matching approaches tend to be computationally inefficient, particularly for large-scale circuits. Deep learning methods have emerged as a potential solution to address these challenges, but existing models fail to efficiently capture global subgraph embeddings or rely on inefficient matching matrices, which limits their effectiveness for large circuits. In this paper, we propose an efficient graph matching approach that utilizes Graph Neural Networks (GNNs) to predict regions of high probability for containing the target circuit. Specifically, we construct various negative samples to enable GNNs to accurately learn the presence of target circuits and develop an approach to directly extracting subgraph embeddings from the entire circuit, which captures global subgraph information and addresses the inefficiency of applying GNNs to all candidate subgraphs. Extensive experiments demonstrate that our approach significantly outperforms existing methods in terms of time efficiency and target region prediction, offering a scalable and effective solution for subgraph matching in large-scale circuits.

Figures

Figures reproduced from arXiv: 2507.19518 by the authors.

Figure 1
Figure 1. Transformation of Circuit Structure into Graph [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The architecture of our proposed method. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Extracting K-hop subgraph embeddings from entire graph [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation studies on the impact of negative samples. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [1]

    Subcircuit identification method based on subgraph isomorphism,

    G. Dong, Y . Zheng, S. He, D. Guo, and L. Li, “Subcircuit identification method based on subgraph isomorphism,” in 2021 IEEE 15th Inter- national Conference on Anti-counterfeiting, Security, and Identification (ASID). IEEE, 2021, pp. 97–101

  2. [2]

    Regds: A reverse engineering framework from gdsii to gate-level netlist,

    R. S. Rajarathnam, Y . Lin, Y . Jin, and D. Z. Pan, “Regds: A reverse engineering framework from gdsii to gate-level netlist,” in 2020 IEEE International Symposium on Hardware Oriented Security and Trust (HOST). IEEE, 2020, pp. 154–163

  3. [3]

    Subgemini: Iden- tifying subcircuits using a fast subgraph isomorphism algorithm,

    M. Ohlrich, C. Ebeling, E. Ginting, and L. Sather, “Subgemini: Iden- tifying subcircuits using a fast subgraph isomorphism algorithm,” in Proceedings of the 30th International Design Automation Conference , 1993, pp. 31–37

  4. [4]

    Subislands: The probabilistic match assignment algorithm for subcircuit recognition,

    N. Rubanov, “Subislands: The probabilistic match assignment algorithm for subcircuit recognition,” IEEE Transactions on Computer-Aided De- sign of Integrated Circuits and Systems , vol. 22, no. 1, pp. 26–38, 2003

  5. [5]

    A (sub) graph isomorphism algorithm for matching large graphs,

    L. P. Cordella, P. Foggia, C. Sansone, and M. Vento, “A (sub) graph isomorphism algorithm for matching large graphs,” IEEE transactions on pattern analysis and machine intelligence , vol. 26, no. 10, pp. 1367– 1372, 2004

  6. [6]

    Ispd 2015 benchmarks with fence regions and routing blockages for detailed- routing-driven placement,

    I. S. Bustany, D. Chinnery, J. R. Shinnerl, and V . Yutsis, “Ispd 2015 benchmarks with fence regions and routing blockages for detailed- routing-driven placement,” in Proceedings of the 2015 Symposium on International Symposium on Physical Design , 2015, pp. 157–164

  7. [7]

    Versatile multi-stage graph neural network for circuit representation,

    S. Yang, Z. Yang, D. Li, Y . Zhang, Z. Zhang, G. Song, and J. Hao, “Versatile multi-stage graph neural network for circuit representation,” Advances in Neural Information Processing Systems , vol. 35, pp. 20 313–20 324, 2022

  8. [8]

    Cktgnn: Circuit graph neural network for electronic design automation,

    Z. Dong, W. Cao, M. Zhang, D. Tao, Y . Chen, and X. Zhang, “Cktgnn: Circuit graph neural network for electronic design automation,” arXiv preprint arXiv:2308.16406, 2023

Show all 33 references
  1. [9]

    Simgnn: A neural network approach to fast graph similarity computation,

    Y . Bai, H. Ding, S. Bian, T. Chen, Y . Sun, and W. Wang, “Simgnn: A neural network approach to fast graph similarity computation,” in Proceedings of the twelfth ACM international conference on web search and data mining , 2019, pp. 384–392

  2. [10]

    Interpretable neural subgraph matching for graph retrieval,

    I. Roy, V . S. B. R. Velugoti, S. Chakrabarti, and A. De, “Interpretable neural subgraph matching for graph retrieval,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 7, 2022, pp. 8115– 8123

  3. [11]

    Neural subgraph matching,

    Z. Lou, J. You, C. Wen, A. Canedo, J. Leskovec et al., “Neural subgraph matching,” arXiv preprint arXiv:2007.03092 , 2020

  4. [12]

    An algorithm for subgraph isomorphism,

    J. R. Ullmann, “An algorithm for subgraph isomorphism,” ACM, vol. 23, no. 1, pp. 31–42, 1976

  5. [13]

    Graph-based technologies for intelligence analysis,

    C. Thayne, S. Greenblatt, and S. Marcus, “Graph-based technologies for intelligence analysis,” ACM, vol. 47, no. 3, pp. 45–47, 2004

  6. [14]

    Template based semantic similarity for security applications,

    B. Aleman-Meza, C. Halaschek-Wiener, S. Sanket Sahoo, A. Sheth, and I. B. Arpinar, “Template based semantic similarity for security applications,” in International Conference on Intelligence and Security Informatics, 2005

  7. [15]

    Substructure similarity search in graph databases,

    X. Yan, P. S. Yu, and J. Han, “Substructure similarity search in graph databases,” in Proceedings of the ACM SIGMOD international conference on Management of data , 2005

  8. [16]

    Graph matching al- gorithms for business process model similarity search,

    R. Dijkman, M. Dumas, and L. Garcia-Banuelos, “Graph matching al- gorithms for business process model similarity search,” in International Conference on Business Process Management , 2005, pp. 48–63

  9. [17]

    Algorithmics and applications of tree and graph searching,

    D. Shasha, J. T. Wang, and R. Giugno, “Algorithmics and applications of tree and graph searching,” in Proceedings of the twenty-first ACM SIGMOD-SGACT symposium on Principles of database systems , 2002, pp. 39–52

  10. [18]

    Graph indexing: a frequent structure-based approach,

    X. Yan, P. S.Yu, and J. Han, “Graph indexing: a frequent structure-based approach,” in Proceedings of the 2004 ACM SIGMOD international conference on Management of data , 2004, pp. 335–346

  11. [19]

    Chemical similarity search- ing,

    P. Willett, J. Baranard, and G. M. Downs, “Chemical similarity search- ing,” Journal of chemical information and computer sciences , vol. 38, no. 6, pp. 983–996, 1998

  12. [20]

    Rascal: Calculation of graph similarity using maximum common edge subgraphs,

    J. W. Raymond, E. J. Gradiner, and P. Willett, “Rascal: Calculation of graph similarity using maximum common edge subgraphs,” The Computer Journal, vol. 45, no. 6, pp. 631–644, 2002

  13. [21]

    Approximately counting triangles in sublinear time,

    T. Eden, A. Levi, D. Ron, and C. Seshadhri, “Approximately counting triangles in sublinear time,” in FOCS, 2015

  14. [22]

    Fast counting of triangles in large real networks without counting: Algorithms and laws,

    C. E. Tsourakakis, “Fast counting of triangles in large real networks without counting: Algorithms and laws,” in ICDM, 2022, pp. 608–617

  15. [23]

    A high-performance subcircuit recognition method based on the nonlinear graph optimization,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfar- dini, “A high-performance subcircuit recognition method based on the nonlinear graph optimization,” IEEE Transactions on Neural Networks , vol. 20, no. 1, pp. 61–80, 2009

  16. [24]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

  17. [25]

    Graph matching networks for learning the similarity of graph structured objects,

    Y . Li, C. Gu, T. Dullien, O. Vinyals, and P. Kohli, “Graph matching networks for learning the similarity of graph structured objects,” in International conference on machine learning. PMLR, 2019, pp. 3835– 3845

  18. [26]

    A high-performance subcircuit recognition method based on the nonlinear graph optimization,

    N. Rubanov, “A high-performance subcircuit recognition method based on the nonlinear graph optimization,” IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems , vol. 25, no. 11, pp. 2353–2363, 2006

  19. [27]

    Gana: Graph convolutional network based automated netlist annotation for analog circuits,

    K. Kunal, T. Dhar, M. Madhusudan, J. Poojary, A. Sharma, W. Xu, S. M. Burns, J. Hu, R. Harjani, and S. S. Sapatnekar, “Gana: Graph convolutional network based automated netlist annotation for analog circuits,” in 2020 Design, Automation & Test in Europe Conference & Exhibition...

  20. [28]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” in The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15 . Springer, ...

  21. [29]

    Neural subgraph isomorphism counting,

    X. Liu, H. Pan, M. He, Y . Song, X. Jiang, and L. Shang, “Neural subgraph isomorphism counting,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp. 1959–1969

  22. [30]

    Challenging the time complexity of exact subgraph isomorphism for huge and dense graphs with vf3,

    V . Carletti, P. Foggia, A. Saggese, and M. Vento, “Challenging the time complexity of exact subgraph isomorphism for huge and dense graphs with vf3,” IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 4, pp. 804–818, 2017

  23. [31]

    A fast projected fixed-point algorithm for large graph matching,

    Y . Lu, K. Huang, and C.-L. Liu, “A fast projected fixed-point algorithm for large graph matching,” Pattern Recognition, vol. 60, pp. 971–982, 2016

  24. [32]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” arXiv preprint arXiv:1810.00826 , 2018

  25. [33]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.