Pith. sign in

REVIEW 4 major objections 6 minor 58 references

CLGNN: A Contrastive Learning-based GNN Model for Betweenness Centrality Prediction on Temporal Graphs

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CLGNN predicts temporal betweenness centrality with up to 31.4x lower MAE than static GNNs and up to 5.7x lower than temporal GNNs, while running up to 663.7x faster than exact computation.

desk verdict A plausible new model for temporal betweenness prediction whose reported gains are undermined by an underspecified evaluation protocol; worth reviewing carefully, not desk-rejecting. read the letter →

arxiv 2506.14122 v1 pith:DZOY24FW submitted 2025-06-17 cs.LG cs.AI

Tianming Zhang , Renbo Zhang , Zhengyi Yang , Yunjun Gao , Bin Cao , Jing Fan This is my paper
classification cs.LGcs.AI
keywords Temporalbetweennesscentralitygraphneuralnetworkscontrastivelearninggraphsclassimbalancepredictioninductivepathcount
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 tries to establish that temporal betweenness centrality (TBC) — a node-importance score measuring how often a node lies on optimal time-respecting paths — can be predicted accurately and cheaply by a graph neural network, despite the fact that in real temporal networks over 95% of nodes have zero TBC. It argues that this extreme imbalance, not model capacity, is the main obstacle: standard GNNs collapse to predicting zero for almost every node and miss the rare high-centrality hubs. To fix this, CLGNN builds an instance graph that preserves path validity and temporal order, encodes structural and temporal features jointly, and adds a clustering-guided contrastive loss that separates zero-, median-, and high-centrality nodes in representation space. The paper reports that on 12 real temporal networks CLGNN outperforms static and temporal GNN baselines by large margins and is up to 663.7 times faster than exact TBC computation. If correct, this would make temporal centrality feasible on large dynamic networks and give practitioners a way to identify genuinely central nodes in streaming interaction data.

What carries the argument

The load-bearing objects are the temporal path-count encoding and the clustering-guided contrastive loss. For each directed interaction $(u,v,t)$, the count $P(u,v,t)=\sum_{t_v \in T_{\text{out}}(v)} I(t_v - t)$ records how many outgoing temporal edges from $v$ can continue the path within the time budget; this scalar, log-transformed and MLP-encoded, tells the network which messages are on valid time-respecting paths and gates message passing. The second mechanism, KContrastNet, first selects the number of clusters $\hat{k}$ by minimizing bootstrap K-means instability, then builds positive pairs from same-cluster nodes with similar TBC and negative pairs from same-cluster nodes with very different TBC, with weights $\beta_{uv}$ inversely or directly proportional to TBC difference. This forces the representation space to separate zero-, median-, and high-centrality nodes while keeping the regression module stable.

What would settle it

Train CLGNN and its baselines on a fully specified held-out protocol — the 50 training networks named, the 12 evaluation networks never seen in training, and learning rates selected on a validation split — then check whether CLGNN's MAE advantage over DBGNN and TGAT persists at the reported 5.7x and 9.4x margins; if the margins shrink to near parity under that protocol, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that pairwise contrastive supervision over cluster-structured node embeddings, guided by the TBC distribution itself, lets a temporal GNN regress TBC values without being swamped by the zero-value majority. Concretely, CLGNN encodes each interaction with a message that concatenates previous-layer node states, a continuous-time encoding, and a temporal path count $P(u,v,t)$ that counts how many valid temporal continuations exist after edge $(u,v,t)$; a dual aggregator combines mean neighbor messages with edge-to-node multi-head attention. A stability-based clustering module estimates the number of TBC-consistent clusters via bootstrap resampling and K-means, and the KContrastNet loss pulls together nodes in the same cluster with similar TBC while pushing apart nodes with very different TBC, reweighted so harder pairs receive larger gradients. ValueNet then maps the final embeddings to TBC values. The paper's evidence is that this pipeline, trained on small graphs, transfers inductively to unseen graphs and beats both static GNNs (up to 31.4x lower MAE, 16.7x higher Spearman) and temporal GNNs (up to 5.7x lower MAE, 3.9x higher Spearman), while running orders of magnitude faster than the exact ETBC algorithm.

Load-bearing premise

The central claim assumes that the reported accuracy gains reflect genuine inductive generalization: the paper says it trains on 50 real temporal networks but never names those networks, describes the train/test split for the 12 evaluation datasets, or clarifies whether the 'best mean score' across three learning rates is chosen on validation or test performance.

Editorial extensions

If this is right

  • Temporal betweenness centrality can be approximated on graphs large enough that exact computation is infeasible, with reported speedups of 4.8x to 663.7x over the exact ETBC algorithm.
  • Important hubs in temporal networks can be identified rather than collapsed into the zero-centrality majority, because the contrastive module explicitly separates mid- and high-value nodes and improves Spearman ranking quality.
  • The model works under multiple optimal path definitions, including shortest, shortest-plus-earliest-arrival, and shortest-plus-latest-departure-plus-earliest-arrival, so one network can serve different temporal semantics.
  • Because the model is inductive, a model trained on small temporal graphs can be applied to unseen graphs, avoiding the per-graph retraining that transductive baselines such as DBGNN require.
  • Contrastive learning, rather than oversampling, undersampling, Tweedie loss, or random graph augmentation, handles the extreme imbalance of centrality distributions on graphs.

Reading between the lines

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

  • Beyond the paper, the same recipe — stability-selected clusters plus TBC-weighted contrastive pairs — should transfer to other imbalanced graph regression targets such as temporal closeness or Katz centrality, because the mechanism only assumes a skewed scalar label per node and an embedding space.
  • The path-count gate $P(u,v,t)$ is a cheap proxy for temporal reachability; extending it to multi-step reachability counts could sharpen the signal on graphs with long time-respecting paths, a testable variation.
  • If the 50 training networks and the exact train/test protocol are disclosed, the inductive claim can be directly verified by held-out evaluation; until then the cross-graph generalization claim is conditional on that protocol.
  • The ablation results suggest that random graph augmentation hurts TBC prediction because it breaks time-respecting paths; a principled next step would be augmentation that edits only causally irrelevant edges.
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 CLGNN, a contrastive-learning GNN for predicting temporal betweenness centrality (TBC). The model constructs an instance graph, augments node representations with temporal path counts and time encodings, uses a dual mean/attention aggregator, and couples a clustering-guided contrastive module (KContrastNet) with a regression head (ValueNet). The authors claim large accuracy gains over static and temporal GNN baselines and large speedups over exact TBC computation, based on 12 real datasets and 30 runs per experiment. The manuscript also includes ablation studies, hyperparameter analyses, and several theoretical appendices.

Significance. If the empirical claims are sound, the paper makes a useful contribution: it is an inductive, scalable TBC regressor rather than a transductive one, and the contrastive imbalance-handling design is a plausible remedy for the extreme skew in TBC distributions. The presentation of 30-run means with standard deviations, comparisons against the exact ETBC method, the fairness intuitions in Appendix F, and the multiple optimal-path-definition study are positive features. However, the headline numbers currently rest on an underspecified evaluation protocol: the 50 training graphs are unnamed, the train/test split is undefined, and learning-rate selection is reported as selecting the 'best mean score' without stating that a validation set was used. These issues must be resolved before the claimed improvements can be accepted.

major comments (4)
  1. [Section 5 (Datasets) and Appendix H] The paper claims that CLGNN is 'trained on small graphs and generalizes to unseen graphs', but the 50 real-world training networks are never named and no train/test split is described. To support the inductive-generalization claim, the authors must list the 50 training networks, explicitly state whether any of the 12 evaluation datasets in Table 4 overlap with them, and describe how the trained model is applied to each test graph.
  2. [Appendix H (Experiment Setup)] The paper states that for each deep learning model it tried three learning rates and 'reported the best mean score', but it does not state that this selection was performed on a validation split. If the best mean is selected by evaluating on the test set, the improvements in Tables 1 and 2 are systematically optimistic. The conclusion only says the learning rate was tuned via grid search. The authors must specify the selection protocol and, if selection was not validation-based, rerun the experiments with a proper validation split.
  3. [Section 5 and Appendix H] The experimental protocol does not describe how the 50-graph training corpus is used: is one model trained on all 50 graphs and then evaluated on each test graph, or is a separate model trained per test dataset with possible fine-tuning? It is also unclear how the transductive DBGNN baseline is adapted to the cross-graph setting. Without this information, the reported head-to-head gains cannot be attributed to inductive generalization rather than to different training setups.
  4. [Section 3 (Definition 2) and Section 5] TBC depends on the temporal-path parameter delta and on the chosen optimal-path definition, but the main experiments (Tables 1 and 2) do not state which definition and which delta value were used to compute ground-truth TBC. Appendix K shows that performance is highly sensitive to the path definition (e.g., sp_hypertext MAE ranges from 6.09 to 50.15), so the main tables are not reproducible without this information.
minor comments (6)
  1. [Abstract and Section 5 (RQ1)] The abstract reports up to 16.7x higher Spearman correlation, while the RQ1 discussion says 16.5x; these numbers should be harmonized.
  2. [Figure 1] The caption contains 'theia-reality-call'; this should be 'ia-reality-call'.
  3. [Conclusion] The text says 'real-world datesets'; this should be 'datasets'.
  4. [Appendix G] Proposition G.1 refers to 'Eq. (xx) of the main paper'; this placeholder should be resolved.
  5. [Appendix K] The phrase 'n contrast' should be 'In contrast'.
  6. [Section 4.1] The temporal path count P(u,v,t) is defined only via the number of later outgoing edges from v, not via actual path enumeration; the paper should explicitly acknowledge this is a heuristic proxy and discuss its limitations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CLGNN is a supervised regressor whose TBC labels enter only through training losses and contrastive pair selection; the reported evaluation-protocol caveats are reliability concerns, not constructional circularity.

full rationale

CLGNN is a supervised regressor: ground-truth TBC values are used only in the training objective, through the regression loss L_regress and through the TBC-difference-based positive/negative pair construction and reweighting in KContrastNet; at inference ValueNet maps learned embeddings to TBC. No equation in Section 4 defines the prediction as a function of the labels or as a fitted constant, and the temporal path-count feature P(u,v,t) is a local reachability signal, not the normalized all-pairs TBC target, so the input is not the output under another name. The comparison baselines (GCN, DrBC, GNN-Bet, TGAT, TATKC, DBGNN, IIOIwD) are external systems, and the self-citations to TATKC [52], ETBC [53], and IIOIwD [54] serve as baseline methods, exact-computation ground truth, or speedup comparators rather than as uniqueness theorems or unverified premises that force the central claim. The main caveats—an unnamed 50-network training set with no disclosed split versus the 12 evaluation datasets, and the statement 'we repeated all experiments for each deep learning model using three learning rates, 0.1, 0.01, and 0.001, and reported the best mean score' (Appendix H) without an explicit validation split—are evaluation-reliability threats that could inflate the headline gains, but they do not make any prediction equivalent to its inputs by construction. Hence no significant circularity.

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

The central claim depends on several hand-set hyperparameters (alpha, lambda, gamma thresholds, tau) and on the modeling assumption that the outgoing-edge count P(u,v,t) approximates path validity. No new physical or structural entities are introduced.

free parameters (5)
  • alpha (loss trade-off) = 0.2 (default)
    Controls balance between contrastive and regression loss; tuned on haggle and used for all datasets.
  • lambda (aggregation weight) = 0.4 (default)
    Blends mean and attention aggregation; tuned on haggle.
  • gamma_pos (positive threshold) = 0.5
    Threshold for positive contrastive pairs relative to median TBC; set by hand, no sensitivity analysis.
  • gamma_neg (negative threshold) = 0.5
    Threshold for negative contrastive pairs; set by hand.
  • tau (temperature) = not specified
    Temperature in InfoNCE loss; not reported in the paper.
assumptions (3)
  • domain assumption The 12 evaluation datasets are representative of real-world temporal graphs and the exact TBC labels computed by ETBC are correct.
    The performance numbers rely on the accuracy of the TBC ground truth from [53].
  • standard math The bootstrap stability clustering estimate is consistent and the i.i.d. resampling assumption holds.
    Appendices D and E provide consistency under the weak law of large numbers; this is a standard statistical argument.
  • ad hoc to paper The temporal path count P(u,v,t), defined as the number of outgoing edges from v with a later timestamp, is a sufficient proxy for path validity.
    This is a modeling choice introduced in Section 4.1 that is not derived from the TBC definition and ignores the time window delta.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLGNN: A Contrastive Learning-based GNN Model for Betweenness Centrality Prediction on Temporal Graphs." pith.science (2026). https://pith.science/paper/DZOY24FW

@misc{pith2026250614122,
  author       = {Pith},
  title        = {Pith review of: CLGNN: A Contrastive Learning-based GNN Model for Betweenness Centrality Prediction on Temporal Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZOY24FW}},
  note         = {Machine review of arXiv:2506.14122}
}
abstract

Temporal Betweenness Centrality (TBC) measures how often a node appears on optimal temporal paths, reflecting its importance in temporal networks. However, exact computation is highly expensive, and real-world TBC distributions are extremely imbalanced. The severe imbalance leads learning-based models to overfit to zero-centrality nodes, resulting in inaccurate TBC predictions and failure to identify truly central nodes. Existing graph neural network (GNN) methods either fail to handle such imbalance or ignore temporal dependencies altogether. To address these issues, we propose a scalable and inductive contrastive learning-based GNN (CLGNN) for accurate and efficient TBC prediction. CLGNN builds an instance graph to preserve path validity and temporal order, then encodes structural and temporal features using dual aggregation, i.e., mean and edge-to-node multi-head attention mechanisms, enhanced by temporal path count and time encodings. A stability-based clustering-guided contrastive module (KContrastNet) is introduced to separate high-, median-, and low-centrality nodes in representation space, mitigating class imbalance, while a regression module (ValueNet) estimates TBC values. CLGNN also supports multiple optimal path definitions to accommodate diverse temporal semantics. Extensive experiments demonstrate the effectiveness and efficiency of CLGNN across diverse benchmarks. CLGNN achieves up to a 663.7~$\times$ speedup compared to state-of-the-art exact TBC computation methods. It outperforms leading static GNN baselines with up to 31.4~$\times$ lower MAE and 16.7~$\times$ higher Spearman correlation, and surpasses state-of-the-art temporal GNNs with up to 5.7~$\times$ lower MAE and 3.9~$\times$ higher Spearman correlation.

Figures

Figures reproduced from arXiv: 2506.14122 by the authors.

Figure 1
Figure 1. Distribution of TBC values in the ia-reality-call dataset temporal paths, which is resource-intensive. To address this, we propose a novel graph neural network (GNN) model designed to accurately predict Temporal Betweenness Centrality (TBC) values. Existing graph learning models primarily focus on either computing betweenness centrality (BC) for static graphs [10, 24] or performing vertex ranking on temporal graphs … view at source ↗
Figure 2
Figure 2. Overview of the Proposed Model: CLGNN [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the representation module 13 [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: depicts the prediction module, which includes a regression head to estimate the temporal betweenness centrality (TBC) values, as well as a contrastive learning module designed to enhance representation separability under extreme imbalance. These components jointly opti…
Figure 5
Figure 5. Figure 5: Model efficiency: comparison of CLGNN and ETBC [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Hyperparameter tuning results over (α, λ) space and individual parameter ablations on the dataset haggle. (a)(b): joint performance surfaces for MAE and Spearman; (c)(d): 1D sensitivity analysis with fixed λ and α, respectively. Note The coefficient α controls the trad…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 56 canonical work pages

  1. [1]

    Borassi and E

    M. Borassi and E. Natale. KADABRA is an adaptive algorithm for betweenness via random approximation. ACM J. Exp. Algorithmics, 24(1):1.2:1–1.2:35, 2019

    work page 2019

  2. [2]

    U. Brandes. A faster algorithm for betweenness centrality.Journal of mathematical sociology, 25(2):163– 177, 2001

    work page 2001

  3. [3]

    Brandes and C

    U. Brandes and C. Pich. Centrality estimation in large networks.Int. J. Bifurc. Chaos, 17(7):2303–2318, 2007

    work page 2007

  4. [4]

    Brody, U

    S. Brody, U. Alon, and E. Yahav. How attentive are graph attention networks? InICLR, 2022

    work page 2022

  5. [5]

    S. Buß, H. Molter, R. Niedermeier, and M. Rymar. Algorithmic aspects of temporal betweenness. InKDD, pages 2084–2092, 2020

    work page 2020

  6. [6]

    Chaintreau, P

    A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass, and J. Scott. Impact of human mobility on opportunistic forwarding algorithms.IEEE Transactions on Mobile Computing, 6(6):606–620, 2007

    work page 2007

  7. [7]

    Cousins, C

    C. Cousins, C. Wohlgemuth, and M. Riondato. Bavarian: Betweenness centrality approximation with variance-aware rademacher averages. InSIGKDD, pages 196–206, 2021

    work page 2021

  8. [8]

    Eagle and A

    N. Eagle and A. Pentland. Reality mining: sensing complex social systems.Personal and Ubiquitous Computing, 10(4):255–268, 2006

    work page 2006

Show all 58 references
  1. [9]

    Erdös, V

    D. Erdös, V . Ishakian, A. Bestavros, and E. Terzi. A divide-and-conquer algorithm for betweenness centrality. InSIAM, pages 433–441, 2015

    2015

  2. [10]

    C. Fan, L. Zeng, Y . Ding, M. Chen, Y . Sun, and Z. Liu. Learning to identify high betweenness centrality nodes from scratch: A novel graph neural network approach. InCIKM, pages 559–568, 2019

    2019

  3. [11]

    Fang and J

    Y . Fang and J. Wang. Selection of the number of clusters via the bootstrap method.Computational Statistics & Data Analysis, 56(3):468–477, 2012

    2012

  4. [12]

    Fournet and A

    J. Fournet and A. Barrat. Contact patterns among high school students.PLoS ONE, 9(9):e107878, 2014. Data collected in collaboration with SocioPatterns. Seehttp://www.sociopatterns.orgfor details

    2014

  5. [13]

    Génois and A

    M. Génois and A. Barrat. Can co-location be used as a proxy for face-to-face contacts?EPJ Data Science, 7(1):1–18, 2018

    2018

  6. [14]

    J. A. Hartigan and M. A. Wong. Algorithm as 136: A k-means clustering algorithm.Journal of the royal statistical society. series c (applied statistics), 28(1):100–108, 1979

    1979

  7. [15]

    Heeg and I

    F. Heeg and I. Scholtes. Using time-aware graph neural networks to predict temporal centralities in dynamic graphs. InNeurIPS, 2024

    2024

  8. [16]

    Isella, J

    L. Isella, J. Stehlé, A. Barrat, C. Cattuto, J. Pinton, and W. Van den Broeck. What’s in a crowd? analysis of face-to-face behavioral networks.Journal of Theoretical Biology, 271(1):166–180, 2011

    2011

  9. [17]

    Jacob, D

    R. Jacob, D. Koschützki, K. A. Lehmann, L. Peeters, and D. Tenfelde-Podehl. Algorithms for centrality indices. InNetwork Analysis, pages 62–82, 2004

    2004

  10. [18]

    T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks, 2017

    2017

  11. [19]

    Kourtellis, T

    N. Kourtellis, T. Alahakoon, R. Simha, A. Iamnitchi, and R. Tripathi. Identifying high betweenness centrality nodes in large social networks.CoRR, abs/1702.06087, 2017

    2017 arXiv

  12. [20]

    J. Kunegis. Konect: the koblenz network collection. InProceedings of the 22nd International Conference on World Wide Web, pages 1343–1350. ACM, 2013

    2013

  13. [21]

    Leskovec and A

    J. Leskovec and A. Krevl. Snap datasets: Stanford large network dataset collection. http://snap. stanford.edu/data, 2014. Accessed: 2025-05-12. 10

    2014

  14. [22]

    Y . Li, Y . Xu, X. Lin, W. Zhang, and Y . Zhang. Ranking on dynamic graphs: An effective and robust band-pass disentangled approach. InProceedings of the ACM on Web Conference 2025, WWW ’25, page 3918–3929, New York, NY , USA, 2025. Association for Computing Machinery

    2025

  15. [23]

    Mastrandrea, J

    R. Mastrandrea, J. Fournet, and A. Barrat. Contact patterns in a high school: a comparison between data collected using wearable sensors, contact diaries and friendship surveys.PLoS ONE, 10(9):e0136497,

  16. [24]

    S. K. Maurya, X. Liu, and T. Murata. Fast approximations of betweenness centrality with graph neural networks. InCIKM, pages 2149–2152, 2019

    2019

  17. [25]

    S. K. Maurya, X. Liu, and T. Murata. Graph neural networks for fast node ranking approximation.ACM Trans. Knowl. Discov. Data, 15(5):78:1–78:32, 2021

    2021

  18. [26]

    M. R. F. Mendonça, A. Barreto, and A. Ziviani. Approximating network centrality measures using node embedding and machine learning.IEEE Trans. Netw. Sci. Eng., 8(1):220–230, 2021

    2021

  19. [27]

    Pareja, G

    A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, T. Schardl, and C. Leiser- son. Evolvegcn: Evolving graph convolutional networks for dynamic graphs. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 5363–5370, 2020

    2020

  20. [28]

    T. P. Peixoto. Netzschleuder: Network catalogue. https://networks.skewed.de/. Accessed: 2025- 05-12

    2025

  21. [29]

    Pellegrina and F

    L. Pellegrina and F. Vandin. SILV AN: estimating betweenness centralities with progressive sampling and non-uniform rademacher bounds.CoRR, abs/2106.03462, 2021

    2021 arXiv

  22. [30]

    Pfeffer and K

    J. Pfeffer and K. M. Carley. k-centralities: local approximations of global measures based on shortest paths. InWWW, pages 1043–1050, 2012

    2012

  23. [31]

    Qarkaxhija, V

    L. Qarkaxhija, V . Perri, and I. Scholtes. De bruijn goes neural: Causality-aware graph neural networks for time series data on dynamic graphs. InLearning on Graphs Conference, pages 51–1. PMLR, 2022

    2022

  24. [32]

    S. C. Regunta, S. H. Tondomker, K. Shukla, and K. Kothapalli. Efficient parallel algorithms for dynamic closeness- and betweenness centrality.Concurr. Comput. Pract. Exp., 35(17), 2023

    2023

  25. [33]

    Riondato and E

    M. Riondato and E. M. Kornaropoulos. Fast approximation of betweenness centrality through sampling. Data Min. Knowl. Discov., 30(2):438–475, 2016

    2016

  26. [34]

    Riondato and E

    M. Riondato and E. Upfal. ABRA: approximating betweenness centrality in static and dynamic graphs with rademacher averages. InSIGKDD, pages 1145–1154, 2016

    2016

  27. [35]

    Rossi, B

    E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. M. Bronstein. Temporal graph networks for deep learning on dynamic graphs.CoRR, abs/2006.10637, 2020

    2006 arXiv

  28. [36]

    R. A. Rossi and N. K. Ahmed. The network data repository with interactive graph analytics and visualiza- tion, 2015. Accessed: May 2025

    2015

  29. [37]

    P. J. Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis.Journal of computational and applied mathematics, 20:53–65, 1987

    1987

  30. [38]

    Sankar, Y

    A. Sankar, Y . Wu, L. Gou, W. Zhang, and H. Yang. Dysat: Deep neural representation learning on dynamic graphs via self-attention networks. InWSDM, pages 519–527, 2020

    2020

  31. [39]

    Santoro and I

    D. Santoro and I. Sarpe. ONBRA: rigorous estimation of the temporal betweenness centrality in temporal networks. InWWW, pages 1579–1588, 2022

    2022

  32. [40]

    A. E. Sariyüce, K. Kaya, E. Saule, and Ü. V . Çatalyürek. Graph manipulations for fast centrality computa- tion.ACM Trans. Knowl. Discov. Data, 11(3):26:1–26:25, 2017

    2017

  33. [41]

    Saunshi, O

    N. Saunshi, O. Plevrakis, S. Arora, M. Khodak, and H. Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. InInternational Conference on Machine Learning, pages 5628–5637. PMLR, 2019

    2019

  34. [42]

    R. L. Thorndike. Who belongs in the family?Psychometrika, 18(4):267–276, 1953

    1953

  35. [43]

    Tsalouchidou, R

    I. Tsalouchidou, R. Baeza-Yates, F. Bonchi, K. Liao, and T. Sellis. Temporal betweenness centrality in dynamic graphs.Int. J. Data Sci. Anal., 9(3):257–272, 2020

    2020

  36. [44]

    van den Oord, Y

    A. van den Oord, Y . Li, and O. Vinyals. Representation learning with contrastive predictive coding.CoRR, abs/1807.03748, 2018

    2018 arXiv

  37. [45]

    Vanhems, A

    P. Vanhems, A. Barrat, C. Cattuto, J.-F. Pinton, N. Khanafer, C. Régis, B.-A. Kim, B. Comte, and N. V oirin. Estimating potential infection transmission routes in hospital wards using wearable proximity sensors. PloS one, 8(9):e73970, 2013

    2013

  38. [46]

    S. Wang, W. Lin, Y . Yang, X. Xiao, and S. Zhou. Efficient route planning on public transportation networks: A labelling approach. InSIGMOD, pages 967–982, 2015. 11

    2015

  39. [47]

    Wang and P

    T. Wang and P. Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. InInternational conference on machine learning, pages 9929–9939. PMLR, 2020

    2020

  40. [48]

    X. Wang, Y . Wang, X. Lin, J. X. Yu, H. Gao, X. Cheng, and D. Yu. Efficient betweenness centrality computation over large heterogeneous information networks.Proc. VLDB Endow., 17(11):3360–3372, 2024

    2024

  41. [49]

    D. Xu, C. Ruan, E. Körpeoglu, S. Kumar, and K. Achan. Inductive representation learning on temporal graphs. InICLR, 2020

    2020

  42. [50]

    Y . Yang, K. Zha, Y . Chen, H. Wang, and D. Katabi. Delving into deep imbalanced regression. InICML, 2021

    2021

  43. [51]

    Zhang, R

    Q. Zhang, R. Li, M. Pan, Y . Dai, G. Wang, and Y . Yuan. Efficient top-k ego-betweenness search. InICDE, pages 380–392, 2022

    2022

  44. [52]

    Zhang, J

    T. Zhang, J. Fang, Z. Yang, B. Cao, and J. Fan. TATKC: A temporal graph neural network for fast approximate temporal katz centrality ranking. InWWW, pages 527–538, 2024

    2024

  45. [53]

    Zhang, Y

    T. Zhang, Y . Gao, J. Zhao, L. Chen, L. Jin, Z. Yang, B. Cao, and J. Fan. Efficient exact and approximate betweenness centrality computation for temporal graphs. InWWW, pages 2395–2406, 2024

    2024

  46. [54]

    Zhang, C

    T. Zhang, C. Hou, R. Jiang, X. Zhang, C. Zhou, K. Tang, and H. Lv. Label informed contrastive pretraining for node importance estimation on knowledge graphs.IEEE Transactions on Neural Networks and Learning Systems, 36(3):4462–4476, 2025

    2025

  47. [55]

    …𝒗𝟐 𝒗𝟑 𝒗𝟒 𝑡&𝑡'𝑡( 𝑡)𝑡* 𝑡*𝑡+ [𝑡

    Y . Zou, T. Li, and Z. Luo. Node centrality approximation for large networks based on inductive graph neural networks.CoRR, abs/2403.04977, 2024. 12 A Details of the CLGNN——Representation Module As shown in Figure 3, the representation module integrates structural and temporal...

    2024 arXiv

  48. [57]

    The positive-pair weightβ uv is strictly decreasing with respect to∆ uv

  49. [58]

    Proof.• Forβ uv = T BCmedian ·γ pos ∆uv , we have: dβuv d∆uv =− T BCmedian ·γ pos ∆2uv <0

    The negative-pair weightβ uw is strictly increasing with respect to∆ uw. Proof.• Forβ uv = T BCmedian ·γ pos ∆uv , we have: dβuv d∆uv =− T BCmedian ·γ pos ∆2uv <0. Hence, βuv decreases as the TBC difference increases, emphasizing the similarity of closely matched pairs. • Forβ...

    2011

  50. [2015]

    See http://www.sociopatterns.org for details

    Data collected in collaboration with SocioPatterns. See http://www.sociopatterns.org for details

Pith tools

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