Pith. sign in

REVIEW 4 major objections 4 minor 28 references

Methodology for Identifying Social Groups within a Transactional Graph

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

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

Pith's one-line read The paper's central claim is that the full set of subgraphs of a given social-group type can be recovered from a transactional graph using only a handful of labeled samples.

desk verdict A clean problem statement and a genuinely fresh reverse-pruning idea, but no experiments, a control-flow bug in Algorithm 1, and an output format in Algorithm 2 that cannot represent the overlapping SGIs the paper itself defines. read the letter →

arxiv 2502.07694 v1 pith:V6NKEFBA submitted 2025-02-11 cs.SI

Maxence Morin , Baptiste Hemery , Fabrice Jeanne , Estelle Pawlowski-Cherrier This is my paper · ORCID
classification cs.SI
keywords SocialNetworkAnalysisTransactionalGraphsSubGraphofInterestTruthIdentificationCommunityDetectionMatchingPruningFeatureVectors
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 framework for recovering all social groups of a given type from a transactional graph, using only a small set of known example groups. It introduces the notion of a SubGraph of Interest (SGI), the subgraph formed by the transactions among members of one group, and formalizes the goal as reconstructing the full set of SGIs of a type from a handful of labeled samples. Two method families are laid out: a generate-and-select pipeline that produces candidate subgraphs and keeps those close to the samples, and a prune-and-connect pipeline that removes nodes and edges far from the samples and reads the remaining connected components as groups. The paper also adapts precision and recall to score recovered groups against ground truth. If the framework works, an organization could discover all groups of a target type—families, colleague circles, fraud rings—from transaction records with only a handful of confirmed examples; the paper does not yet report experimental results.

What carries the argument

The load-bearing object is the SubGraph of Interest (SGI), defined as the subgraph of observed transactions made by one Group of Interest. The mechanism that carries the argument is feature-vector comparison: every candidate subgraph, node, or edge is mapped to a fixed-size vector (via hand-engineered metrics or learned embeddings), and the cosine distance $d_c(A,B)=A\cdot B/(\|A\|\|B\|)$ to the feature vectors of the training samples is compared to a threshold $\Gamma$. Elements below the threshold are kept; elements above it are discarded or pruned. In the pruning variant the key operator is the graph-pruning function $\nabla$ with four modes, and the final connected components are declared to be the recovered SGIs.

What would settle it

Construct a synthetic transactional graph whose ground-truth SGIs contain two structurally distinct variants of the same group type, train only on samples from one variant, and run the framework; if recall on the held-out variant drops sharply while precision stays high, the assumption that the samples represent the whole type has failed and the central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that for one group type, the whole set $S$ of SubGraphs of Interest can be recovered from only the small sample set $S_n$, provided each candidate subgraph is represented by a fixed-size feature vector and compared by cosine distance to features of the samples. In the first approach, candidate subgraphs come from overlapping community detection or from graph matching against a query built as the Maximum Common Subgraph of the samples; in the second, nodes and edges whose features are too distant from sample elements are pruned, and the connected components of the pruned graph are the output. The paper argues that this recovers SGIs that community detection would miss because SGI transactions need not be dense, and that graph matching alone would miss because topology alone cannot distinguish types that differ only in context. Thus the contribution is a generic framework with three method variants and an evaluation scheme, rather than a demonstrated result.

Load-bearing premise

The framework assumes that the few labeled example groups capture all structural and contextual variation of that group type, so cosine distance to these samples reliably separates group members from non-members.

Editorial extensions

If this is right

  • For a fixed group type, an organization would go from a few manually confirmed groups to a candidate list of all groups of that type present in the transaction graph.
  • Because SGIs are allowed to overlap, the recovered groups can place one user in multiple groups, matching reality where the same person belongs to a family and a fraud network.
  • The prune-and-connect variant can in principle separate subgraphs with identical topology if their surrounding context differs, since context enters through node and edge features.
  • The adapted precision, recall, and match functions give a way to judge whether a recovered set $\hat{S}$ is close to the social truth $S$ without requiring exact equality of subgraphs.

Reading between the lines

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

  • Inference: The method's practical ceiling depends on feature quality; if two SGI types share both topology and local features, no amount of threshold tuning will separate them, so the framework implicitly requires the feature set to encode context.
  • Inference: A natural stress test is to shrink $|S_n|$ toward 1 and measure recall; the paper gives no guidance on how few is too few, so this boundary is open.
  • Inference: The pruning variant could be adapted to streaming transaction graphs by updating node and edge feature vectors incrementally, but the paper does not address temporal drift.
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 / 4 minor

Summary. The paper introduces a framework for recovering a set S of 'SubGraphs of Interest' (SGIs) in a transactional multigraph from a small sample set Sn of known SGIs. It formalizes the problem in Section 2, proposes two approaches in Section 3 — a FirstApproach based on candidate subgraph generation, feature vectors, and thresholded cosine-distance selection, and a SecondApproach based on node/edge feature pruning followed by connected-component extraction — and adapts precision and recall metrics with a match function in Section 4. The paper explicitly states that experimental evaluation is deferred to a separate future article.

Significance. If validated, the framework would address a genuine gap between generic community detection and graph matching by targeting specific group types with context-aware features, and the proposed adaptation of precision/recall via the match function is a reasonable starting point for evaluation. The paper also gives clear definitions of SGIs, GoIs, and the distinction between social and transactional truth. However, the contribution as submitted is a methodological template rather than a demonstrated method: there are no experiments, datasets, or theoretical guarantees supporting the central claim that S can be recovered from Sn. Moreover, the connected-component output of Algorithm 2 is structurally incompatible with the overlapping SGIs allowed in Section 2.2, and Algorithm 1 contains a control-flow bug that limits its output to at most one subgraph. These issues currently invalidate the two proposed algorithms for the problem as specified.

major comments (4)
  1. [§2.2 and §3.2, Algorithm 2 (lines 22–23)] Section 2.2 explicitly states that 'SGIs can overlap. Thus, a node can belong to several SGIs.' Algorithm 2, however, terminates with 'return connected(G)', i.e., it outputs the connected components of the pruned graph. Connected components are disjoint by definition, so if two SGIs share a node or an edge they fall into the same connected component and are returned as a single merged subgraph. The four proposed pruning variants (∇simple, ∇node, ∇edge, ∇majority) all feed into this same connected-components output. To recover overlapping SGIs, the framework would need a second decomposition step that splits connected components at overlap boundaries, but no such step is defined. This is a structural mismatch between the declared problem and the algorithm, and it directly undermines the central recovery claim.
  2. [§3.1, Algorithm 1 (lines 4–11)] The 'break' statement inside the loop over candidate subgraphs terminates the iteration immediately after the first accepted candidate. As a result, 'results' contains at most one element, so the algorithm cannot output a set Ŝ approximating the full set S when multiple SGIs exist. The intended behavior is presumably to continue scanning all candidates and collect every subgraph that passes the check; as written, the algorithm contradicts the Section 2.2 objective of recovering the entire set S.
  3. [§1 and §5] The paper's stated second contribution is to 'Evaluate this framework,' yet the manuscript contains no experiments, datasets, or benchmarks. Section 5 states that 'Experiments using the framework are in progress and will be presented in a separate article.' Consequently, the central claim that the framework recovers the entire set S from a small sample set Sn is entirely unsupported by empirical evidence or by a formal guarantee. Without such support, the paper does not establish its main thesis.
  4. [§3.1, §3.2, and §4] The feature functions are only sketched as non-exhaustive lists ('This list is non-exhaustive'), and all thresholds (Γ, Γ_node, Γ_edge, Γ_extra, Γ_missing, Γ_size) are left as context-dependent free parameters. The methods are therefore not fully specified as reproducible algorithms; different instantiations could produce very different outputs. This is a missing-support issue for the claimed general framework, and it would need to be addressed either by concrete default choices justified by experiments or by a formal sensitivity analysis.
minor comments (4)
  1. [§2.2] The notation 'we use S and S instead of Stype and Stype' is confusing because the two sets are not typographically distinguished in the text; please use clear calligraphic or bold variants consistently.
  2. [§3.2, Algorithm 2 line 22] Algorithm 2 calls ∇(G, Vbad, Ebad) before the four concrete pruning functions are defined; please specify which variant is intended or parameterize the call.
  3. [§4, Eq. (15)–(16)] The precision and recall definitions count each predicted subgraph once, but when SGIs overlap a single predicted component may match multiple ground-truth subgraphs; the matching assignment should be clarified to avoid double counting in the overlap setting.
  4. [Figure 3] The figure caption and surrounding text should explicitly state which nodes and edges are part of Sn, which are part of S, and which are distractors, as the current description is terse.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is an inductive, sample-based heuristic whose claims are unverified, but its outputs are not definitionally tied to its inputs.

full rationale

The paper's derivation chain is: define SGIs and the target set S from a postulated social truth; propose two methods that use the sample set Sn to select or prune candidate subgraphs via cosine distance on feature vectors; and define precision/recall-style evaluation against S. No load-bearing step equates an output with an input by construction. In Algorithm 1, 'check' returns true when a candidate's feature vector is within threshold Gamma of any sample, which is a nearest-neighbor selection rule, not a hidden identity: the paper does not define 'SGI of type T' as 'subgraph close to Sn', so selecting by sample similarity is a heuristic, not a tautology. In Algorithm 2, the paper explicitly states 'We suppose that the resulting connected components are all SGIs in S', which is an unproven modeling assumption rather than a definitional reduction; the thresholds Gamma_node and Gamma_edge are free parameters, not quantities fitted to S or to Sn in a way that would force the result. The evaluation section compares outputs to S using thresholds that are 'set according to the context', which is a limitation on reproducibility, not circularity. There are no self-citations in the load-bearing references; the cited subgraph feature and embedding methods are external. The paper's own admission in Section 5 that 'Experiments using the framework are in progress and will be presented in a separate article' means the central recovery claim is unsupported, but unsupported claims are a correctness/evidence risk, not circularity. The connected-components versus overlapping-SGI mismatch flagged by the skeptic is likewise an internal correctness gap in Algorithm 2, not a reduction of the output to the input. Accordingly, no circular step is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The framework introduces two new conceptual entities and relies on several uncalibrated thresholds and domain assumptions. No experiments, datasets, or code are provided, so the ledger cannot be closed empirically.

free parameters (4)
  • Cosine distance threshold Γ (first approach) = Not set
    Upper bound on dc(Fκ,Fψ) in Algorithm 1's check function; introduced as a tunable constant with no value or calibration specified.
  • Node and edge thresholds Γ_node, Γ_edge (second approach) = Not set
    Maximum cosine distance allowed from a node or edge feature to training samples in Algorithm 2; values are left to context.
  • Majority edge ratio 0.5 in ∇majority = 0.5
    Hand-chosen threshold in Eq. (7) deciding which bad nodes are spared; no justification is provided.
  • Evaluation thresholds Γ_extra, Γ_missing, Γ_size = Not set
    Used in Eqs. (9)-(11) to define when a predicted subgraph matches a true one; the paper says they are set according to context.
assumptions (4)
  • domain assumption Edges in the transactional graph are traces of underlying social relationships, and all relationships will induce a transaction given enough time.
    Stated in the introduction: 'Transactions are traces of underlying social relationships... If we extend the period of time, every social relationship may induce a transaction.' The entire framing depends on this link.
  • domain assumption Feature vectors computed from the observable graph are sufficient to distinguish nodes, edges, and subgraphs inside an SGI from those outside it.
    Algorithms 1 and 2 compare feature vectors via cosine distance to samples. No experiment or theoretical argument establishes this discriminative power.
  • domain assumption The small sample set Sn is representative of the full set S for the target group type.
    Section 2.2 acknowledges Sn has very few samples, but the framework uses these samples as the only training signal to recover all SGIs of that type.
  • standard math Graph isomorphism and graph edit distance complexity results are correct, as standard results.
    The paper cites NP-completeness of isomorphism and NP-hardness of edit distance; these are background results, not central to the framework.
invented entities (2)
  • Group of Interest (GoI) and SubGraph of Interest (SGI)
    purpose: Formalize the target social groups and their observable transaction patterns, allowing overlap between groups.
    Introduced in Sections 1 and 2.2. The concepts are defined by the paper, but no external dataset or measurement demonstrates they correspond to recoverable structure.
  • Social truth and transactional truth
    purpose: Separate the complete but inaccessible set of social groups from the observed transaction graph.
    Theoretical constructs used to state the problem; they are not measured or validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Methodology for Identifying Social Groups within a Transactional Graph." pith.science (2026). https://pith.science/paper/V6NKEFBA

@misc{pith2026250207694,
  author       = {Pith},
  title        = {Pith review of: Methodology for Identifying Social Groups within a Transactional Graph},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6NKEFBA}},
  note         = {Machine review of arXiv:2502.07694}
}
read the original abstract

Social network analysis is pivotal for organizations aiming to leverage the vast amounts of data generated from user interactions on social media and other digital platforms. These interactions often reveal complex social structures, such as tightly-knit groups based on common interests, which are crucial for enhancing service personalization or fraud detection. Traditional methods like community detection and graph matching, while useful, often fall short of accurately identifying specific groups of users. This paper introduces a novel framework specifically designed to identify groups of users within transactional graphs by focusing on the contextual and structural nuances that define these groups.

Figures

Figures reproduced from arXiv: 2502.07694 by the authors.

Figure 1
Figure 1. Differences between the observed transactional trut [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The first two subgraphs represent two different types o [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. An example of a multigraph. Dashed nodes are from a sub [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 19 canonical work pages

  1. [1]

    : Sub2Vec: Feature Learning for Subgraphs

    Adhikari, B., Zhang, Y., Ramakrishnan, N., Prakash, B.A. : Sub2Vec: Feature Learning for Subgraphs. In: Phung, D., Tseng, V.S., Webb, G. I., Ho, B., Ganji, M., Rashidi, L. (eds.) Advances in Knowledge Discovery and D ata Mining. pp. 170–182. Lecture Notes in Computer Science, Springer Inter national Publishing (2018). https://doi.org/10.1007/978-3-319-93037-4_14

    doi:10.1007/978-3-319-93037-4_14 2018

  2. [2]

    Subgraph Neural Networks

    Alsentzer, E., Finlayson, S., Li, M., Zitnik, M.: Subgrap h Neural Networks. In: Advances in Neural Information Processing Systems. vol. 33 , pp. 8017–8029 (2020). https://doi.org/10.48550/ARXIV.2006.10538

    work page Pith review arXiv doi:10.48550/arxiv.2006.10538 2020

  3. [3]

    In: Proceedings of the Forty-Eighth Annual ACM Symposium on The ory of Comput- ing

    Babai, L.: Graph isomorphism in quasipolynomial time [ex tended abstract]. In: Proceedings of the Forty-Eighth Annual ACM Symposium on The ory of Comput- ing. p. 684–697. STOC ’16, Association for Computing Machin ery, New York, NY, USA (2016). https://doi.org/10.1145/2897518.2897542

    work page arXiv 2016

  4. [4]

    Physics Re ports 486(3–5), 75–174 (Feb 2010)

    Fortunato, S.: Community detection in graphs. Physics Re ports 486(3–5), 75–174 (Feb 2010). https://doi.org/10.1016/j.physrep.2009.11.002

    doi:10.1016/j.physrep.2009.11.002 2010

  5. [5]

    In: AAAI Fall Symposium: Capturing and Using Patt erns for Evidence Detection (2006), https://api.semanticscholar.org/Cor pusID:12917214

    Gallagher, B.: Matching structure and semantics: A surve y on graph-based pattern matching. In: AAAI Fall Symposium: Capturing and Using Patt erns for Evidence Detection (2006), https://api.semanticscholar.org/Cor pusID:12917214

    work page 2006

  6. [6]

    Proceedings of the National Academy of Sciences 99(12), 7821–7826 (Jun 2002)

    Girvan, M., Newman, M.E.J.: Community structure in socia l and biological net- works. Proceedings of the National Academy of Sciences 99(12), 7821–7826 (Jun 2002). https://doi.org/10.1073/pnas.122653799

    doi:10.1073/pnas.122653799 2002

  7. [7]

    In: BCC (2020), https://api.semanticscholar.org/CorpusID: 226237505

    Grohe, M., Neuen, D.: Recent advances on the graph isomorp hism problem. In: BCC (2020), https://api.semanticscholar.org/CorpusID: 226237505

    work page 2020

  8. [8]

    In: Proceedings of the 22nd ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining

    Grover, A., Leskovec, J.: Node2vec: Scalable Feature Lea rning for Net- works. In: Proceedings of the 22nd ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining. ACM (Aug 201 6). https://doi.org/10.1145/2939672.2939754

    work page arXiv

Show all 28 references
  1. [9]

    Machine Learning 110(3), 451–456 (Mar 2021)

    Hand, D.J., Christen, P., Kirielle, N.: F*: an interpreta ble transforma- tion of the f-measure. Machine Learning 110(3), 451–456 (Mar 2021). https://doi.org/10.1007/s10994-021-05964-1

    2021 doi

  2. [10]

    John Wiley & Sons , Inc., USA (1962)

    Iverson, K.E.: A programming language. John Wiley & Sons , Inc., USA (1962)

    1962

  3. [11]

    In: International Conference on Lear ning Repre- sentations

    Kipf, T.N., Welling, M.: Semi-Supervised Classificatio n with Graph Con- volutional Networks. In: International Conference on Lear ning Repre- sentations. Poster (2017). https://doi.org/10.48550/AR XIV.1609.02907, https://openreview.net/forum?id=SJU4ayYgl

    2017 doi arXiv

  4. [12]

    Statistical Analysis an d Data Mining 5(4), 265– 283 (Jun 2012)

    Li, G., Semerci, M., Yener, B., Zaki, M.J.: Effective grap h classification based on topological and label attributes. Statistical Analysis an d Data Mining 5(4), 265– 283 (Jun 2012). https://doi.org/10.1002/sam.11153

    2012 doi

  5. [13]

    In: 2021 3rd International Conference on Advances in Comput er Technol- ogy, Information Science and Communication (CTISC)

    Liu, Q., Su, Y., Peng, Q., Chen, K., Lu, Y.: An overlapping commu- nity detection algorithm for label propagation based on nod e influence. In: 2021 3rd International Conference on Advances in Comput er Technol- ogy, Information Science and Communication (CTISC). pp. 18 3–187 (...

    2021

  6. [14]

    In: International Conferenc e on Learning Represen- tations

    Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient Est imation of Word Rep- resentations in Vector Space. In: International Conferenc e on Learning Represen- tations. Poster (2013-01). https://doi.org/10.48550/AR XIV.1301.3781

    2013 doi arXiv

  7. [15]

    In: Proceedings of the 13th International Workshop on Mining and Learning with Graphs (MLG)

    Narayanan, A., Mahinthan, C., Venkatesan, R., Chen, L., Liu, Y., Jaiswal, S.: graph2vec: Learning Distributed Representations of Graph s. In: Proceedings of the 13th International Workshop on Mining and Learning with Graphs (MLG). Workshop on Mining and Learning with Graphs, A...

    2017 doi

  8. [16]

    In: Proceedings of the 20th ACM SIGKDD Internat ional Conference on Knowledge Discovery and Data Mining

    Perozzi, B., Al-Rfou, R., Skiena, S.: DeepWalk: Online L earning of Social Repre- sentations. In: Proceedings of the 20th ACM SIGKDD Internat ional Conference on Knowledge Discovery and Data Mining. pp. 701–710. Associat ion for Computing Machinery (2014). https://doi.org/10.1...

    2014 doi

  9. [17]

    In: Proceedi ngs of the 36th Interna- tional Conference on Neural Information Processing System s

    Roy, I., Chakrabarti, S., De, A.: Maximum common subgrap h guided graph re- trieval: late and early interaction networks. In: Proceedi ngs of the 36th Interna- tional Conference on Neural Information Processing System s. NIPS ’22, Curran Associates Inc., Red Hook, NY, USA (2024)

    2024

  10. [18]

    Inter national Journal of Approxi- mate Reasoning 50(7), 969–978 (2009)

    Salakhutdinov, R., Hinton, G.: Semantic hashing. Inter national Journal of Approxi- mate Reasoning 50(7), 969–978 (2009). https://doi.org/10.1016/j.ijar.20 08.11.006, https://www.sciencedirect.com/science/article/pii/S0888613X08001813, special Section on Graphical Models and I...

    2009 doi

  11. [19]

    In: Proceedings of the 25th ACM SIGKDD Inte rna- tional Conference on Knowledge Discovery and Data Mining

    Shchur, O., Günnemann, S.: Overlapping Community Detec tion with Graph Neural Networks. In: Proceedings of the 25th ACM SIGKDD Inte rna- tional Conference on Knowledge Discovery and Data Mining. D eep Learning on Graphs Workshop (DLG), Association for Computing Machin ery (201...

    2019 doi arXiv

  12. [20]

    Jour nal of King Saud University - Computer and Information Sciences 34(9), 7765–7774 (2022)

    Singh, D., Garg, R.: Ni-louvain: A novel algorithm to de- tect overlapping communities with influence analysis. Jour nal of King Saud University - Computer and Information Sciences 34(9), 7765–7774 (2022). https://doi.org/10.1016/j.jksuc i.2021.07.006, https://www.sciencedirect...

    2022 doi

  13. [21]

    In: Sattar, A., Kang, B.h

    Sokolova, M., Japkowicz, N., Szpakowicz, S.: Beyond acc uracy, f-score and roc: A family of discriminant measures for performance evaluatio n. In: Sattar, A., Kang, B.h. (eds.) AI 2006: Advances in Artificial Intelligence. pp . 1015–1021. Springer Berlin Heidelberg, Berlin, Hei...

    2006

  14. [22]

    : SUGAR: Sub- graph Neural Network with Reinforcement Pooling and Self-S upervised Mu- tual Information Mechanism

    Sun, Q., Li, J., Peng, H., Wu, J., Ning, Y., Yu, P.S., He, L. : SUGAR: Sub- graph Neural Network with Reinforcement Pooling and Self-S upervised Mu- tual Information Mechanism. In: Proceedings of the Web Conf erence 2021. pp. 2081–2091. WWW ’21, Association for Computing Machi...

    2021

  15. [23]

    ACM Trans

    Wang, C., Wang, C., Wang, Z., Ye, X., Yu, P.S.: Edge2vec: E dge-based so- cial network embedding. ACM Trans. Knowl. Discov. Data 14(4) (may 2020). https://doi.org/10.1145/3391298

    2020 doi

  16. [24]

    Nauchno-Technicheskaya In formatsia, Series 2(9), 12–16 (1968), https://www.iti.zcu.cz/wl2018/pdf/wl_pa per_translation.pdf

    Weisfeiler, B.Y., Leman, A.A.: The reduction of a graph t o canonical form and the algebra which appears therein. Nauchno-Technicheskaya In formatsia, Series 2(9), 12–16 (1968), https://www.iti.zcu.cz/wl2018/pdf/wl_pa per_translation.pdf

    1968

  17. [25]

    https://doi.org/10.48550/arxiv.1810.00826

    Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful ar e graph neural networks? (2019). https://doi.org/10.48550/arxiv.1810.00826

    2019 doi arXiv

  18. [26]

    In: Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval

    Yan, J., Yin, X.C., Lin, W., Deng, C., Zha, H., Yang, X.: A s hort sur- vey of recent advances in graph matching. In: Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. pp. 1 67–174. ICMR ’16, Association for Computing Machinery, New York, NY, USA...

    2016

  19. [27]

    Zeng, Z., Tung, A.K.H., Wang, J., Feng, J., Zhou, L.: Comp aring stars: on ap- proximating graph edit distance. Proc. VLDB Endow. 2(1), 25–36 (aug 2009). https://doi.org/10.14778/1687627.1687631

    2009

  20. [28]

    Zhu, X., Ghahramani, Z.: Learning from labeled and unlab eled data with label propagation. Tech. rep., Carnegie Mellon Univer sity (2002), https://pages.cs.wisc.edu/~jerryzhu/pub/CMU-CALD-02-107.pdf

    2002

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.