Pith. sign in

REVIEW 6 major objections 6 minor 45 references

RANA: Robust Active Learning for Noisy Network Alignment

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

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

Pith's one-line read The paper claims that its active learning framework RANA outperforms all tested baselines on network alignment under structural and labeling noise, by coupling a noise-aware selection objective with a model-assisted, twin-pair label…

desk verdict RANA tackles a real gap in noisy network alignment, but the core confidence derivation is wrong and the budget comparison is unfair, so the empirical claims are not currently supported. read the letter →

arxiv 2507.22434 v2 pith:ZVHNR27G submitted 2025-07-30 cs.LG

Yixuan Nan , Xixun Lin , Yanmin Shang , Zhuofan Li , Can Zhao , Yanan Cao This is my paper
classification cs.LG
keywords networkalignmentactivelearninglabelnoisestructuralanchorlinkpredictiontwinnodepairsnoise-awareconfidencegraph
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

Network alignment — matching the same user, movie, or protein across two networks — usually relies on supervised models that need pre-labeled anchor links, yet real annotations and network edges are noisy. This paper proposes RANA, an active learning framework that decides which node pairs to query and then cleans the returned labels before they enter training. The paper claims that RANA consistently beats existing active learning and alignment baselines across three real-world datasets and several alignment models, and that its accuracy degrades more slowly than baselines as structural noise or oracle error increases. If that holds, alignment models can be trained with fewer and less trustworthy annotations than current pipelines require.

What carries the argument

The load-bearing mechanism is the noise-aware confidence $C_{ij}$ of Definition 4. It combines the oracle confidence $C_{orc}$, the model prediction confidence $C^m_{ij} = Acc \cdot p_{ij}$, and the cleanliness score $cs_{ij}$, then routes each node pair into one of three labeling regimes: trust the model, trust the oracle, or query a twin node pair to resolve disagreement. Twin node pairs are found by minimizing feature plus two-hop structural similarity distance, and the selection stage maximizes the number of activated nodes, where activation uses the expected-Jacobian influence score $\hat{I}(v_j, v_i, k)$ scaled by node pair confidence and an activation threshold $\theta$. This machinery is what lets RANA prefer informative pairs that are also in structurally reliable neighborhoods and then assign the final label from multiple sources.

What would settle it

On a synthetic alignment task with known ground truth, compare the oracle's label for each twin node pair with the true label; if twin labels are not correct more often than the original oracle labels on equivalent pairs, the Definition 8 denoising gain disappears. Separately, recompute Definition 4's disagreement confidence with the correct normalizer $C^m_{ij}(1-C_{orc}) + (1-C^m_{ij})C_{orc}$ instead of $1-C_{orc}C^m_{ij}$; if the selected node pairs change, the claimed posterior values are not valid.

Watch

Extended reading notes

Core claim

The central discovery is that both kinds of noise can be handled inside the active learning loop rather than after training. For structural noise, RANA computes a cleanliness score for each candidate node pair — the average cosine feature similarity between a node and its neighbors — and uses it to discount pairs sitting in noisy neighborhoods. For labeling noise, RANA splits pairs into confidence regions: when model confidence exceeds oracle accuracy the model labels the pair itself; in the intermediate region, if the oracle and model disagree, RANA queries a twin node pair (the most similar pair by features and two-hop structure) and uses the oracle's twin label to break the tie; in the low region the oracle label is kept. The paper reports that on the Facebook-Twitter dataset this design gains up to 6.24% accuracy over the best baseline, and that the framework improves the performance of FINAL, IsoRank, PALE, and PARROT when wrapped around them.

Load-bearing premise

The method's denoising gain rests on trusting the oracle's label for a twin node pair as reliable evidence about the original pair, as if the twin label were ground truth.

Editorial extensions

If this is right

  • RANA is model-agnostic: the paper shows accuracy improvements when the same active learning loop is wrapped around FINAL, IsoRank, PALE, and PARROT.
  • Under added noisy edges, RANA's accuracy drops only 0.0060 from 0% to 25% noise on the Douban test, while the compared baselines drop more, so the selection module buffers structural noise.
  • As oracle accuracy falls from 0.9 to 0.6 on the Facebook-Twitter dataset, RANA's accuracy declines more slowly than baselines, making it usable with unreliable annotators.
  • When model accuracy exceeds the oracle's, RANA switches to model-assisted labeling, so later iterations can skip oracle queries and reduce annotation cost.

Reading between the lines

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

  • RANA's budget is counted in node pairs, not oracle calls: every disagreement case costs two oracle queries (the original pair plus its twin). A fair cost comparison would plot accuracy against total oracle queries, which would narrow or widen the gap depending on disagreement frequency.
  • The twin-pair denoising rule could generalize to other graph tasks with noisy oracles, such as fraud detection or protein interaction prediction, whenever a second highly similar sample can be identified automatically.
  • The cleanliness score is currently a fixed average of neighbor feature similarities; a testable extension would estimate per-node noise rates from repeated oracle labels and reweight cleanliness accordingly, making the selection module adaptive.
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

6 major / 6 minor

Summary. RANA proposes an active learning framework for network alignment that simultaneously handles structural noise (noisy edges) and labeling noise (noisy oracle). It consists of a Noise-aware Selection Module, which scores candidate node pairs using a combination of oracle accuracy, model prediction confidence, and a graph-based cleanliness score, and a Label Denoising Module, which uses model-assisted labeling for high-confidence pairs and queries 'twin' node pairs to resolve model/oracle disagreements. The paper reports experiments on three dataset pairs using several alignment models (FINAL, PALE, IsoRank, PARROT) and claims consistent accuracy improvements. Appendix A derives the confidence formulas underlying both modules.

Significance. If the proposed mechanism worked as described, RANA would be a useful contribution to active learning on noisy graphs, addressing both label and structural noise within a single framework. The authors provide code, ablations, and experiments across multiple alignment backbones, which are concrete strengths. However, the central denoising mechanism is not supported: the confidence posteriors in Definition 4 and Appendix A contain a mathematical error, the twin-label evidence is not modeled probabilistically, and the budget comparison does not account for twin queries. These issues bear directly on the paper's main claims, so the current version does not establish the proposed advantage.

major comments (6)
  1. [Appendix A, Eqs. (11)–(13); Definition 4] Case 2 and Case 3 of Appendix A use 1 − C_orc·C_m as the probability that model and oracle disagree. For binary labels, disagreement occurs when exactly one of the two is correct, which has probability C_m(1−C_orc)+(1−C_m)C_orc; the event that both are incorrect is an agreement (both produce the same wrong label) and is not part of the disagreement event. Consequently Eqs. (12) and (13) are not conditional posterior probabilities, and the confidence values defined in Definition 4 are not valid posteriors. This is load-bearing because these values drive node pair selection and label trust decisions.
  2. [Definition 7–8; Appendix A Cases 2–3] The twin node pair label is treated as decisive evidence to choose between the model and oracle labels, but the paper provides no model of the twin oracle's accuracy or of the correlation between the original and twin labels. The twin pair is selected as the most similar pair using the same features and structure used for alignment, so its label is not an independent noisy draw; a noisy oracle can be wrong on both pairs in a correlated way. Without an explicit error model (or a differently sourced oracle for the twin), the posterior formulas in Definition 4 and Appendix A cannot justify the denoising gain claimed in Section 5.4 and Table 5.
  3. [Section 5.2, Figure 3(a)] The query budget comparison is unfair: every uncertain query that triggers the twin mechanism costs two oracle calls (the original pair and its twin), but Figure 3(a) plots accuracy against k without counting twin queries. The claim that RANA achieves better accuracy under the same labeling budget is therefore not supported. The budget accounting must include twin queries, or the comparison must be with a total-call budget.
  4. [Eq. (6) and Definition 3] The model-assisted labeling step uses the model's own predictions to label high-confidence pairs when Acc·p > α, and those labels are added to the training set. The next round's Acc is then computed on that same training set, which now contains model-generated labels. This creates a self-training feedback loop: Acc can become inflated by the model's own confident predictions, and the thresholds (α, γ) are not calibrated against a noise-free source. The circularity affects both C_m in Definition 3 and the labeling decisions in Eq. (6), so the reported improvements cannot be attributed to the denoising mechanism alone.
  5. [Eq. (7)] Definition 7 defines the twin node pair as the 'most similar' pair but uses arg min of S(X̂_s_itwin, X̂_s_i)+S(...). If S is a similarity function (cosine similarity is suggested), larger values should indicate greater similarity, so the correct operation is arg max. The notation X̂_s = X_s(A_s)^2 is also dimensionally ambiguous (X_s is N×d while A_s^2 is N×N); presumably the intended feature aggregation is (A_s)^2 X_s or similar. As written, the twin-pair construction is not well-defined, and it is unclear what was actually implemented in the experiments.
  6. [Tables 2–3; Figure 3] All numerical comparisons report means over five runs without standard deviations or significance tests, yet the central claim is that RANA 'consistently outperforms' baselines. Several differences are small (e.g., Table 2, Douban at 0.2: RANA 0.5304 vs. TopMatchings 0.5291), and the reader cannot assess whether the gains are statistically meaningful. Please report per-run values, error bars, and tests (e.g., paired t-test or Wilcoxon) for the main tables and figures.
minor comments (6)
  1. [Definition 2, Eq. (1)] The text claims the cleanliness score is an average feature similarity, but Eq. (1) sums similarities without dividing by node degree. Please make the normalization explicit.
  2. [Section 5.1] The text says six datasets are used; since they are paired into three network-alignment datasets, the wording is misleading.
  3. [Appendix C.1] The Facebook-Twitter dataset is stated to have no node attributes, but the cleanliness score relies on feature similarity. Please explain how the score is computed for attribute-free networks.
  4. [Table 2] Including PARROT and NAME as rows when the alignment model is FINAL is confusing; those are alignment models, not query strategies. Please clarify the baseline protocol.
  5. [Figure 4] The y-axis label 'Acc%1(%)' appears to be a typo for Acc@1(%).
  6. [Algorithm 1] The model accuracy Acc is not defined in the algorithm input; please state how it is estimated at each round.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are benchmarked against external ground truth; the Appendix's twin-oracle posterior error is a correctness issue, not a by-construction equivalence.

full rationale

RANA is an active-learning heuristic: confidence scores (Eqs. 2-4) and label-assignment rules (Eq. 6, Definition 8) are defined from the model's own predicted probabilities, the oracle accuracy, and twin-node similarity. This is self-training in the machine-learning sense, but it is not circular in the sense used here: the paper's headline quantity, alignment accuracy on ground-truth anchor links (Acc@k, MAP), is measured against held-out/external ground truth, not reconstructed from the confidence definitions. The model-assisted labeling loop (Eq. 6 labels by Acc*p>alpha, and the next round uses Acc on the growing set) can reinforce model errors, but that is a known algorithm behavior, not a by-construction equality between prediction and input. The Appendix A derivation of Definition 4 is mathematically defective: for binary independent labels, P(disagreement) is C_m(1-C_orc)+(1-C_m)C_orc, not 1-C_orc*C_m, and the formulas in Cases 2-3 silently treat the twin oracle label as perfectly reliable even though the twin pair is labeled by the same noisy oracle (Section 4.1, Definition 7). That is an invalid and unjustified probabilistic derivation, and a serious correctness risk for the claimed denoising benefit, but it is not a circular reduction: the final label rule and the confidence score are not identical to the evaluation metric, and no load-bearing conclusion is imported from the authors' prior citations. There are no self-citations in the reference list that carry the argument. Accordingly, no circular step can be exhibited under the required standard, and the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

RANA's central mechanism rests on hand-set thresholds (theta, gamma) and on the assumption that twin node pair labels from the same noisy oracle are reliable evidence. The cleanliness score is an invented measure with no external validation and is undefined on the attribute-free Facebook-Twitter dataset. The model-assisted labeling loop makes training accuracy a self-referential confidence measure.

free parameters (5)
  • activation threshold theta = 0.05
    Hand-set in Section 5.1; controls which nodes count as activated in Eq. (4); no sensitivity analysis.
  • minimum acceptable threshold gamma = 0.01
    Hand-set in Section 5.1; defines low-confidence region boundaries in Definitions 4 and 8; no sensitivity analysis.
  • influence propagation steps k
    Definition 5 depends on k but no value is reported in the experiments; missing hyperparameter that affects the activation set.
  • oracle accuracy alpha = 0.8 in budget experiment
    Treated as a known prior in Section 4.1; varied in Figure 3(b) but not reported for Tables 2-3.
  • query budget k = 100 in main tables
    Set per experiment; affects results and interacts with the un-counted twin queries.
assumptions (5)
  • domain assumption Oracle and model prediction errors are independent
    Used in Appendix A to multiply C_orc and C^m_ij; not justified and likely false because both make errors on the same hard node pairs.
  • ad hoc to paper Twin node pair label from the noisy oracle is reliable evidence for the original pair
    Definitions 7-8 and Appendix A Cases 2-3 treat twin label agreement as deciding which source (model or oracle) is correct, without modeling twin oracle noise or correlation.
  • domain assumption Training accuracy Acc is a valid estimate of model reliability on unlabeled node pairs
    Eq. (2) and Eq. (6) use Acc as model confidence; training accuracy can be inflated and does not measure generalization to unlabeled pairs.
  • domain assumption Cosine similarity between node attributes is defined for all datasets
    Definition 2 uses node features; Facebook-Twitter has no node attributes (Appendix C.1), so the cleanliness score is undefined unless an unstated substitute is used.
  • standard math The coverage objective is submodular and the greedy marginal gain is valid
    Cites Nemhauser for greedy submodular maximization; the specific F in Eq. (5) is a coverage function, but the paper does not verify all conditions for the finite-horizon noisy setting.
invented entities (2)
  • cleanliness score
    purpose: Measures local structural reliability to down-weight noisy neighborhoods during node pair selection
    Introduced in Definition 2; no external benchmark verifies that neighbor attribute similarity is a valid proxy for edge noise; undefined on attribute-free datasets.
  • twin node pairs
    purpose: Provides an extra noisy oracle label to resolve disagreement between model and oracle during labeling
    Introduced in Definition 7; treated as a reliable tie-breaker without evidence that twin pairs have higher label accuracy, and the extra oracle query is not budgeted in experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RANA: Robust Active Learning for Noisy Network Alignment." pith.science (2026). https://pith.science/paper/ZVHNR27G

@misc{pith2026250722434,
  author       = {Pith},
  title        = {Pith review of: RANA: Robust Active Learning for Noisy Network Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZVHNR27G}},
  note         = {Machine review of arXiv:2507.22434}
}
read the original abstract

Network alignment has attracted widespread attention in various fields. However, most existing works mainly focus on the problem of label sparsity, while overlooking the issue of noise in network alignment, which can substantially undermine model performance. Such noise mainly includes structural noise from noisy edges and labeling noise caused by human-induced and process-driven errors. To address these problems, we propose RANA, a Robust Active learning framework for noisy Network Alignment. RANA effectively tackles both structure noise and label noise while addressing the sparsity of anchor link annotations, which can improve the robustness of network alignment models. Specifically, RANA introduces the proposed Noise-aware Selection Module and the Label Denoising Module to address structural noise and labeling noise, respectively. In the first module, we design a noise-aware maximization objective to select node pairs, incorporating a cleanliness score to address structural noise. In the second module, we propose a novel multi-source fusion denoising strategy that leverages model and twin node pairs labeling to provide more accurate labels for node pairs. Empirical results on three real-world datasets demonstrate that RANA outperforms state-of-the-art active learning-based methods in alignment accuracy. Our code is available at https://github.com/YXNan0110/RANA.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [1]

    Aral and P

    S. Aral and P. S. Dhillon. Social influence maximization under empiri- cal influence models.Nature human behaviour, 2(6):375–382, 2018

    work page 2018

  2. [2]

    Bartunov, A

    S. Bartunov, A. Korshunov, S.-T. Park, W. Ryu, and H. Lee. Joint link- attribute user identity resolution in online social networks. InProceed- ings of the 6th International Conference on Knowledge Discovery and Data Mining, Workshop on Social Network Mining and Analysis. ACM, pages 12–16, 2012

    work page 2012

  3. [3]

    H. Cai, V . W. Zheng, and K. C.-C. Chang. Active learning for graph embedding.arXiv preprint arXiv:1705.05085, 2017

    arXiv 2017

  4. [4]

    Cao and Y

    X. Cao and Y . Yu. Asnets: A benchmark dataset of aligned social net- works for cross-platform user modeling. InProceedings of the 25th ACM International on Conference on Information and Knowledge Man- agement, pages 1881–1884, 2016

    work page 2016

  5. [5]

    Deep Active Learning for Anchor User Prediction

    A. Cheng, C. Zhou, H. Yang, J. Wu, L. Li, J. Tan, and L. Guo. Deep active learning for anchor user prediction.arXiv preprint arXiv:1906.07318, 2019

    work page Pith review arXiv 1906

  6. [6]

    Clark and J

    C. Clark and J. Kalita. A multiobjective memetic algorithm for ppi network alignment.Bioinformatics, 31(12):1988–1998, 2015

    work page 1988

  7. [7]

    Z. Deng, J. Sang, and C. Xu. Personalized video recommendation based on cross-platform user modeling. In2013 IEEE International Confer- ence on Multimedia and Expo (ICME), pages 1–6. IEEE, 2013

    work page 2013

  8. [8]

    F. E. Faisal, H. Zhao, and T. Milenkovi ´c. Global network alignment in the context of aging. InProceedings of the 5th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics, pages 580–580, 2014

    work page 2014

Show all 45 references
  1. [9]

    Fuchsgruber, T

    D. Fuchsgruber, T. Wollschläger, B. Charpentier, A. Oroz, and S. Gün- nemann. Uncertainty for active learning on graphs.arXiv preprint arXiv:2405.01462, 2024

    2024 arXiv

  2. [10]

    Glaria, C

    F. Glaria, C. Hernández, S. Ladra, G. Navarro, and L. Salinas. Compact structure for sparse undirected graphs based on a clique graph partition. Information Sciences, 544:485–499, 2021

    2021

  3. [11]

    Huang, G

    S. Huang, G. Lee, Z. Bao, and S. Pan. Cost-effective data labelling for graph neural networks. InProceedings of the ACM Web Conference 2024, pages 353–364, 2024

    2024

  4. [12]

    T. T. Huynh, C. T. Duong, T. T. Nguyen, V . T. Van, A. Sattar, H. Yin, and Q. V . H. Nguyen. Network alignment with holistic embeddings.IEEE Transactions on Knowledge and Data Engineering, 35(2):1881–1894, 2021

    2021

  5. [13]

    Jiang, P

    M. Jiang, P. Cui, X. Chen, F. Wang, W. Zhu, and S. Yang. Social rec- ommendation with cross-domain transferable knowledge.IEEE trans- actions on knowledge and data engineering, 27(11):3084–3097, 2015

    2015

  6. [14]

    A. J. Joshi, F. Porikli, and N. Papanikolopoulos. Multi-class active learning for image classification. In2009 ieee conference on computer vision and pattern recognition, pages 2372–2379. IEEE, 2009

    2009

  7. [15]

    Z. Kang, H. Pan, S. C. Hoi, and Z. Xu. Robust graph learning from noisy data.IEEE transactions on cybernetics, 50(5):1833–1843, 2019

    2019

  8. [16]

    Konyushkova, R

    K. Konyushkova, R. Sznitman, and P. Fua. Learning active learning from data.Advances in neural information processing systems, 30, 2017

    2017

  9. [17]

    Koutra, H

    D. Koutra, H. Tong, and D. Lubensky. Big-align: Fast bipartite graph alignment. In2013 IEEE 13th international conference on data mining, pages 389–398. IEEE, 2013

    2013

  10. [18]

    Y . Li, J. Yin, and L. Chen. Seal: Semisupervised adversarial active learning on attributed graphs.IEEE Transactions on Neural Networks and Learning Systems, 32(7):3136–3147, 2020

    2020

  11. [19]

    Y . Li, J. Ren, J. Liu, and Y . Chang. Deep sparse autoencoder predic- tion model based on adversarial learning for cross-domain recommen- dations.Knowledge-Based Systems, 220:106948, 2021

    2021

  12. [20]

    Malmi, A

    E. Malmi, A. Gionis, and E. Terzi. Active network alignment: A matching-based approach. InProceedings of the 2017 ACM on Confer- ence on Information and Knowledge Management, pages 1687–1696, 2017

    2017

  13. [21]

    T. Man, H. Shen, S. Liu, X. Jin, and X. Cheng. Predict anchor links across social networks via an embedding approach. InIjcai, volume 16, pages 1823–1829, 2016

    2016

  14. [22]

    Maskey and Y .-R

    S. Maskey and Y .-R. Cho. Survey of biological network alignment: Cross-species analysis of conserved systems. In2019 IEEE Interna- tional Conference on Bioinformatics and Biomedicine (BIBM), pages 2090–2096. IEEE, 2019

    2019

  15. [23]

    G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. An analysis of ap- proximations for maximizing submodular set functions—i.Mathemat- ical programming, 14:265–294, 1978

    1978

  16. [24]

    Y . Ren, C. C. Aggarwal, and J. Zhang. Activeiter: Meta diagram based active learning in social networks alignment.IEEE Transactions on Knowledge and Data Engineering, 33(5):1848–1860, 2019

    2019

  17. [25]

    Rossi and N

    R. Rossi and N. Ahmed. The network data repository with interactive graph analytics and visualization. InProceedings of the AAAI confer- ence on artificial intelligence, volume 29, 2015

    2015

  18. [26]

    Saxena and J

    S. Saxena and J. Chandra. A survey on network alignment: approaches, applications and future directions. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 8216– 8224, 2024

    2024

  19. [27]

    Senette, M

    C. Senette, M. Siino, and M. Tesconi. User identity linkage on social networks: A review of modern techniques and applications.IEEE Ac- cess, 2024

    2024

  20. [28]

    Singh, J

    R. Singh, J. Xu, and B. Berger. Pairwise global alignment of protein interaction networks by matching neighborhood topology. InAnnual international conference on research in computational molecular biol- ogy, pages 16–31. Springer, 2007

    2007

  21. [29]

    Singh, J

    R. Singh, J. Xu, and B. Berger. Global alignment of multiple pro- tein interaction networks with application to functional orthology de- tection.Proceedings of the National Academy of Sciences, 105(35): 12763–12768, 2008

    2008

  22. [30]

    S. Tan, Z. Guan, D. Cai, X. Qin, J. Bu, and C. Chen. Mapping users across networks by manifold alignment on hypergraph. InProceedings of the AAAI Conference on Artificial Intelligence, volume 28, 2014

    2014

  23. [31]

    H. T. Trung, N. T. Toan, T. Van Vinh, H. T. Dat, D. C. Thang, N. Q. V . Hung, and A. Sattar. A comparative study on network alignment tech- niques.Expert Systems with Applications, 140:112883, 2020

    2020

  24. [32]

    H. T. Trung, T. Van Vinh, N. T. Tam, H. Yin, M. Weidlich, and N. Q. V . Hung. Adaptive network alignment with unsupervised and multi-order convolutional networks. In2020 IEEE 36th International Conference on Data Engineering (ICDE), pages 85–96. IEEE, 2020

    2020

  25. [33]

    Van Vlasselaer, C

    V . Van Vlasselaer, C. Bravo, O. Caelen, T. Eliassi-Rad, L. Akoglu, M. Snoeck, and B. Baesens. Apate: A novel approach for automated credit card transaction fraud detection using network-based extensions. Decision support systems, 75:38–48, 2015

    2015

  26. [34]

    Wang and J

    H. Wang and J. Leskovec. Unifying graph convolutional neural net- works and label propagation.arXiv preprint arXiv:2002.06755, 2020

    2002 arXiv

  27. [35]

    Wu and Y

    Y . Wu and Y . Yuan. Robust offline active learning on graphs.arXiv preprint arXiv:2408.07941, 2024

    2024 arXiv

  28. [36]

    H. Yin, Q. Wang, K. Zheng, Z. Li, J. Yang, and X. Zhou. Social influence-based group representation learning for group recommenda- tion. In2019 IEEE 35th International Conference on Data Engineering (ICDE), pages 566–577. IEEE, 2019

    2019

  29. [37]

    Zafarani and H

    R. Zafarani and H. Liu. Connecting corresponding identities across communities. InProceedings of the international AAAI conference on web and social media, volume 3, pages 354–357, 2009

    2009

  30. [38]

    Z. Zeng, S. Zhang, Y . Xia, and H. Tong. Parrot: Position-aware regu- larized optimal transport for network alignment. InProceedings of the ACM Web Conference 2023, pages 372–382, 2023

    2023

  31. [39]

    Zhang and Y

    A.-p. Zhang and Y . Chen. A real-time detection algorithm for abnor- mal users in multi relationship social networks based on deep neural network. InInternational Conference on Advanced Hybrid Information Processing, pages 179–190. Springer, 2021

    2021

  32. [40]

    Zhang and H

    S. Zhang and H. Tong. Final: Fast attributed network alignment. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1345–1354, 2016

    2016

  33. [41]

    Zhang, Y

    W. Zhang, Y . Wang, Z. You, M. Cao, P. Huang, J. Shan, Z. Yang, and B. Cui. Rim: Reliable influence-based active learning on graphs.Ad- vances in Neural Information Processing Systems, 34:27978–27990, 2021

    2021

  34. [42]

    Zhang, Y

    W. Zhang, Y . Wang, Z. You, M. Cao, P. Huang, J. Shan, Z. Yang, and B. Cui. Information gain propagation: a new way to graph active learn- ing with soft labels.arXiv preprint arXiv:2203.01093, 2022

    2022 arXiv

  35. [43]

    Zhang, K

    Y . Zhang, K. Sharma, and Y . Liu. Capturing cross-platform inter- action for identifying coordinated accounts of misinformation cam- paigns. InEuropean Conference on Information Retrieval, pages 694–

  36. [44]

    Y . Zhao, C. Li, J. Peng, X. Fang, F. Huang, S. Wang, X. Xie, and J. Gong. Beyond the overlapping users: Cross-domain recommendation via adaptive anchor link learning. InProceedings of the 46th interna- tional ACM SIGIR conference on research and development in informa- tion r...

    2023

  37. [45]

    Q. Zhou, L. Li, X. Wu, N. Cao, L. Ying, and H. Tong. Attent: Active attributed network alignment. InProceedings of the Web Conference 2021, pages 3896–3906, 2021. A Proof of Confidence Calculation Next, we will demonstrate how the confidence of node pairs is calcu- lated. When...

    2021

Pith tools

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