Pith. sign in

REVIEW 3 major objections 6 minor 58 references

Bridging Source and Target Domains via Link Prediction for Unsupervised Domain Adaptation on Graphs

T0 review · 3 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 Predicting and inserting cross-domain edges lets graph message passing pull target nodes into the source distribution, reducing domain shift even under label mismatch.

desk verdict A genuinely new input-level graph UDA idea with solid empirical gains, but the label-shift claim is not verified on the tasks where it matters most. read the letter →

arxiv 2505.24055 v1 pith:PV7QKSC6 submitted 2025-05-29 cs.LG cs.AI

Yilong Wang , Tianxiang Zhao , Zongyu Wu , Suhang Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords graphneuralnetworksunsuperviseddomainadaptationcross-domainlinkpredictionmessagepassinglabelshiftnodeclassificationongraphsmutualinformationloss
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

This paper sets out to show that unsupervised domain adaptation on graphs can be done at the input level instead of only aligning final embeddings: predict and insert cross-domain edges between source and target graphs, and let GNN message passing do the alignment. The framework trains a link predictor on the source graph's own edges, uses it to connect each target node to similar source nodes with learned edge weights, and adds a mutual-information loss to keep target nodes from losing their discriminative identity. On citation, blog, and Twitch graphs, the paper reports consistent accuracy and AUROC gains over prior adversarial and discrepancy-based graph domain adaptation baselines, including settings where the class distribution differs across domains. The central claim is that these bridge edges align class-conditional embedding distributions, not just marginal distributions, and that is why the method tolerates label shift.

What carries the argument

The load-bearing object is the cross-domain link predictor and the edge-insertion update on the combined graph: a shared-weight GNN encoder produces node embeddings, an MLP link predictor scores candidate cross-domain edges, and edges with predicted probability above a threshold are inserted with adaptive weights. A top-50 similarity candidate set per target node keeps per-epoch edge insertion at $O(n_t)$ after a one-time $O(n_s n_t)$ pre-selection. The inserted edges carry the argument because they turn each target node's neighborhood into a mixture of target and source nodes, so GNN message passing performs the domain alignment directly on the input graph rather than through a separate alignment loss on the encoder's last layer. The mutual-information loss plays a supporting role by contrasting combined-graph and target-graph embeddings of the same node to preserve discriminative information.

What would settle it

Take a target graph with known labels, log the endpoints of every inserted cross-domain edge, and compute the fraction that are intra-class; if the fraction is low and performance still beats the no-bridge baseline, the paper's explanation that intra-class bridges carry the benefit is wrong. More directly, on a synthetic source-target pair constructed so that cross-domain nodes of the same class are less similar than cross-domain nodes of different classes, CMPGNN should lose to a no-edge GCN baseline if the alignment mechanism is doing the work.

Watch

Extended reading notes

Core claim

The central claim is that modifying the target graph itself, by adding cross-domain edges predicted with a learned link predictor, aligns target node embeddings with the source domain more directly than aligning the final layer outputs of an encoder, and that this alignment is class-conditional rather than merely marginal. The link predictor is trained by reconstructing edges inside the source graph, so it learns which embedding pairs plausibly correspond to edges; during adaptation it scores each target node against a pre-selected candidate set of similar source nodes and inserts edges whose predicted probability exceeds a threshold, using the probability as the edge weight. Once these bridges exist, a target node's neighborhood contains labeled source nodes of the same predicted class, so message passing shifts that target node's embedding toward the source class center. A mutual-information loss contrasts each target node's combined-graph embedding with its original target-graph embedding, preventing over-smoothing and the loss of target-specific discriminative information. On the ACM/DBLP, Blog, and Twitch benchmarks, the paper reports that this pipeline outperforms adversarial and discrepancy-based graph domain adaptation baselines and keeps its advantage when the source and target label proportions differ.

Load-bearing premise

The method assumes that a node's class is what makes it look similar to nodes in the other domain, so the automatically inserted bridge edges mostly connect same-class pairs; if feature shift or heterophily makes different-class nodes look equally similar, the inserted edges would propagate wrong-class messages and the claimed label-shift robustness would not follow.

Editorial extensions

If this is right

  • Graph unsupervised domain adaptation can be treated as a graph-editing problem: modifying the input graph itself lets a source-trained GNN classifier generalize without requiring a custom alignment objective.
  • Because the alignment is enforced on the conditional distribution $P(h|y)$ through intra-class bridge edges, the approach should remain applicable when source and target label distributions differ, the setting where marginal-distribution alignment methods break down.
  • The learned edge weights act as differentiable bridges, so the same link-prediction machinery could be reused whenever a model must decide which cross-domain node pairs are safe to exchange information.
  • Any input-level augmentation that pulls domains together must be paired with an identity-preserving regularizer such as the mutual-information loss, or the inserted edges will smooth away class discriminability.

Reading between the lines

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

  • The paper does not test this directly, but its mechanism implies a strong causal prediction: filtering the inserted edges to keep only same-class cross-domain pairs should reproduce the reported gains, while keeping only cross-class pairs should erase or reverse them.
  • A natural extension is multi-source graph domain adaptation, where each target node could build candidate sets from several source graphs and the link predictor's reconstruction loss would need to account for differing source structures.
  • On heterophilic or strongly feature-shifted domains, the framework could be made more robust by down-weighting edges whose source-side class prediction is low-confidence or by decorrelating the link-prediction embeddings from domain-specific signals; neither is explored in the paper.
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

3 major / 6 minor

Summary. The paper proposes CMPGNN, an unsupervised graph domain adaptation method that inserts cross-domain edges into a combined source-target graph. A link predictor is trained to reconstruct observed intra-domain edges and then applied to pre-selected candidate pairs; edges with predicted probability above a threshold are added with learned weights, so that GNN message passing over the augmented graph aligns target-node neighborhoods with the source domain. The training objective combines a source classification loss, a target entropy loss, the link-prediction reconstruction loss, and a contrastive mutual-information loss intended to preserve target-node discriminability. The central claim is that this input-level graph modification aligns conditional distributions and is insensitive to disproportional label distributions. Experiments on ACM/DBLP, Blog, large citation, and Twitch datasets report consistent accuracy/AUROC gains over several graph domain adaptation baselines.

Significance. If the mechanism works as claimed, the paper offers a genuinely different perspective on graph UDA: instead of aligning final-layer embeddings through adversarial or discrepancy objectives, it modifies the input graph so that message passing itself reduces distribution shift. This is a valuable direction and could address a known weakness of marginal alignment under label shift. The empirical package is reasonably complete for a conference submission: multiple benchmarks, fully- and semi-supervised source settings, ablations, hyperparameter sensitivity, and standard deviations over five runs. However, the central label-shift robustness claim is not yet backed by direct evidence where it matters most, and there is a technical tension in the link-predictor training objective that the authors need to resolve.

major comments (3)
  1. [Section 4.2.1, Eq. (3)] The link prediction loss is computed over neighborhoods in the initial combined graph, which is block-diagonal with no cross-domain edges. Every source-target pair is therefore a non-neighbor and can be sampled by the negative sampler in Eq. (3); the text explicitly states that the sample encompasses both intra-domain and cross-domain node pairs. This means the predictor is trained to output probability close to 0 for cross-domain pairs, while Section 4.2.2 later inserts edges only when the predicted probability exceeds t=0.94. This is a train/inference mismatch for the core component of the method. Please either modify the negative sampling so that cross-domain pairs are not treated as negatives, or provide a formal or empirical justification for why probabilities trained to be low on cross-domain pairs are still meaningful at the high threshold. The issue is load-bearing because the entire method depends on these cross-domain edge probabilities.
  2. [Section 4.2.2 and Section 5.5.1] The label-shift robustness claim rests on the premise that top-K similar cross-domain node pairs are predominantly same-class, but intra-class edge precision is measured only for ACM->DBLP and Blog1->Blog2 in Fig. 7, and class-wise center distances are reported only for ACM/DBLP in Fig. 9. The Twitch experiments in Table 2 are the ones described as having significant label shift and are the basis for the abstract's claim of insensitivity to disproportional label distributions. Please report, for at least the DE->EN and FR->EN tasks, the intra-class precision of inserted edges or class-wise conditional distribution distances, using target labels for analysis only, and monitor this precision over training to rule out error accumulation from the jointly trained encoder and link predictor. Without this evidence, the causal story that inserted edges align P(H|Y) rather than merely improving marginal alignment is not directly verified where it is most load-bearing.
  3. [Section 4.2.2 and Algorithm 1] The candidate set is pre-selected using an unspecified 'pre-trained encoder' before training begins, but the pretraining procedure is never described. Since the candidate set is fixed and the link predictor can only insert edges among these candidates, the quality of the pretrained encoder is a critical determinant of the method's success. Please specify the pretraining objective and data, and ideally analyze sensitivity to the candidate size K or update the candidate set periodically during training. As written, this step is not fully reproducible.
minor comments (6)
  1. [Algorithm 1] Line 7 labels the entropy loss as L_{2;5}; this should be L_entropy to match Eq. (8).
  2. [Section 5.4] The text says the hyperparameters govern 'the RA loss L_LP', but Eq. (9) uses L_LP with coefficient lambda_2; the acronym 'RA' appears to be a typo and should be corrected.
  3. [Section 2] The section title 'Relate Work' should be 'Related Work'.
  4. [Abstract] The phrase 'we proposed a novel framework' should be 'we propose' for consistency with the rest of the paper.
  5. [Section 5.2, Table 2] The statement that baseline results in Table 2 are 'inherited from work [20]' should specify whether the evaluation protocol (data splits, preprocessing, and hyperparameter settings) is identical to that of [20], so that the comparison is apples-to-apples.
  6. [Section 3.2] The CSBM preliminary experiment inserts edges based directly on feature similarity with a threshold rather than through the learned link predictor, and reports only t-SNE visualizations; adding a quantitative conditional-shift metric on this synthetic setup would make the motivation more convincing.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild internal loop from same-embedding edge selection and alignment check; no definitional circularity.

  1. other [Section 4.2.2 (Cross-domain Edges Insertion), Eq. (5) and Fig. 9]
    "For each node E8 in the target domain V_t, we compute its embedding similarity of the pre-trained encoder with every node E9 in the source domain V_s. We then select the top-K most similar nodes from the source domain to construct a candidate set C8 for each target node E8. ... if the predicted connection probability s_ij is larger than the threshold t, then we add the new edge (E8,E9) into the modified combined graph."

    The candidate set that defines which cross-domain edges can be inserted is chosen by embedding similarity under the same encoder whose output is later used to measure class-wise alignment (Fig. 9). Since the link predictor's score is s_ij = sigmoid(z_i^T z_j), another dot-product in that same space, the inserted edges are essentially the high-similarity pairs selected by the encoder. Showing that target embeddings move toward source embeddings after aggregating over these self-selected neighbors is partly a restatement of the selection criterion rather than an independent confirmation. However, target labels are used only for evaluation, and the reported accuracy/AUROC is not fitted to target outcomes, so this is a mild internal loop, not full circularity.

full rationale

The paper is self-contained against external benchmarks. Target labels are not used in any training objective: the losses are source cross-entropy, target entropy, link reconstruction on observed edges, and the mutual-information loss; target labels enter only at final evaluation. The link predictor is trained on observed intra-domain edges, and cross-domain edges are extrapolated, so the method is not fitting target outcomes. The label-shift robustness claim rests on the empirical homophily assumption of Section 4.2.2, which is only partially verified (Fig. 7 covers ACM to DBLP and Blog1 to Blog2, not the Twitch label-shift tasks), but that is a correctness/evidence gap, not a circular derivation. Self-citations in the bibliography (e.g., references [54] and [55]) are not load-bearing in the argument. The only mild circularity is the internal loop described above, where the edge-selection criterion and the alignment evaluation share the same embedding space; this does not undermine the independent classification results.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on one data-geometry premise (similar embeddings imply same class), one transfer premise for the link predictor, and one effect premise for message passing. There are no new physical or latent entities; the free parameters are loss weights and fixed thresholds that affect the method's behavior and are not derived from theory.

free parameters (4)
  • lambda_1, lambda_2, lambda_3 = not reported; grid searched over {1e-3, 1e-1, 1, 10, 100}
    Weights for entropy, link prediction, and mutual information losses in Eq. (9); tuned per dataset, but exact chosen values are not given in the main text.
  • edge insertion threshold t = 0.94
    Fixed threshold for adding a predicted cross-domain edge, chosen by hand in Section 4.2.2.
  • candidate set size K = 50
    Number of top-similar source nodes kept per target node; fixed for all datasets in Section 4.2.2.
  • MI temperature tau = 0.3
    Temperature in the mutual information loss in Eq. (6); fixed in experiments.
assumptions (3)
  • domain assumption Cross-domain node pairs with similar embeddings are predominantly same-class pairs, so top-K similarity candidates contain enough intra-class edges for message passing to help.
    Used in Section 4.2.2 to justify edge candidate selection and in Figure 1; if heterophily or feature shift breaks this, inserted edges add noise.
  • domain assumption A link predictor trained by reconstructing intra-domain edges (Eq. 3) transfers to unseen cross-domain edge existence.
    Section 4.2.1 uses source/combined-graph reconstruction as a surrogate for cross-domain links; no cross-domain edge labels exist.
  • ad hoc to paper Message passing over inserted edges reduces conditional shift P(h|y) enough to improve target classification under label shift.
    Asserted in Sections 1 and 3.2 based on a CSBM toy example; no theorem or bound is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Source and Target Domains via Link Prediction for Unsupervised Domain Adaptation on Graphs." pith.science (2026). https://pith.science/paper/PV7QKSC6

@misc{pith2026250524055,
  author       = {Pith},
  title        = {Pith review of: Bridging Source and Target Domains via Link Prediction for Unsupervised Domain Adaptation on Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PV7QKSC6}},
  note         = {Machine review of arXiv:2505.24055}
}
read the original abstract

Graph neural networks (GNNs) have shown great ability for node classification on graphs. However, the success of GNNs relies on abundant labeled data, while obtaining high-quality labels is costly and challenging, especially for newly emerging domains. Hence, unsupervised domain adaptation (UDA), which trains a classifier on the labeled source graph and adapts it to the unlabeled target graph, is attracting increasing attention. Various approaches have been proposed to alleviate the distribution shift between the source and target graphs to facilitate the classifier adaptation. However, most of them simply adopt existing UDA techniques developed for independent and identically distributed data to gain domain-invariant node embeddings for graphs, which do not fully consider the graph structure and message-passing mechanism of GNNs during the adaptation and will fail when label distribution shift exists among domains. In this paper, we proposed a novel framework that adopts link prediction to connect nodes between source and target graphs, which can facilitate message-passing between the source and target graphs and augment the target nodes to have ``in-distribution'' neighborhoods with the source domain. This strategy modified the target graph on the input level to reduce its deviation from the source domain in the embedding space and is insensitive to disproportional label distributions across domains. To prevent the loss of discriminative information in the target graph, we further design a novel identity-preserving learning objective, which guides the learning of the edge insertion module together with reconstruction and adaptation losses. Experimental results on real-world datasets demonstrate the effectiveness of our framework.

Figures

Figures reproduced from arXiv: 2505.24055 by the authors.

Figure 1
Figure 1. This !gure depicts a simpli!ed cross-domain graph for node classi!cation, with pale blue lines indicating poten￾tial cross-domain edges. class distribution, and will fail when the disproportional label dis￾tribution occurs among domains. A detailed analysis of this issue is provided in Section 3.2. (iii) These methods do not fully lever￾age the unique properties of graph structures and the intrinsic message-passing … view at source ↗
Figure 2
Figure 2. Visualization of embeddings on the generated [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of embeddings on the generated [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Ablation Study Results [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Hyperparameter Sensitivity. performance of Random Link is similar to that of GCN+ DA, under￾scoring the critical role of the link predictor. This !nding suggests that discerning informative cross-domain connections is critical for the success of CMPGNN; and (ii) The fu…
Figure 8
Figure 8. Figure 8: Changes in embedding similarity during training. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Embedding Center Distance Among Domains. The [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 49 canonical work pages

  1. [1]

    James Atwood and Don Towsley. 2016. Di￿usion-Convolutional Neural Networks. In Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems, NIPS 2016 . 1993–2001

    work page 2016

  2. [2]

    Wendong Bi, Xueqi Cheng, Bingbing Xu, Xiaoqian Sun, Li Xu, and Huawei Shen. 2023. Bridged-gnn: Knowledge bridge learning for e￿ective knowledge transfer. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management . 99–109

    work page 2023

  3. [3]

    Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. 2014. Spectral Networks and Locally Connected Networks on Graphs. In 2nd International Conference on Learning Representations, ICLR 2014

    work page 2014

  4. [4]

    Quanyu Dai, Xiao-Ming Wu, Jiaren Xiao, Xiao Shen, and Dan Wang. 2023. Graph Transfer Learning via Adversarial Domain Adaptation With Graph Convolution. IEEE Trans. Knowl. Data Eng. 35 (2023), 4908–4922

    work page 2023

  5. [5]

    Yash Deshpande, Subhabrata Sen, Andrea Montanari, and Elchanan Mossel. 2018. Contextual Stochastic Block Models. InAnnual Conference on Thirty-second Neural Information Processing Systems, NeurIPS 2018 . 8590–8602

    work page 2018

  6. [6]

    Matthias Fey and Jan Eric Lenssen. 2019. Fast graph representation learning with PyTorch Geometric. arXiv preprint arXiv:1903.02428 (2019)

    arXiv 2019

  7. [7]

    Lempitsky

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor S. Lempitsky. 2016. Domain-Adversarial Training of Neural Networks. J. Mach. Learn. Res. 17 (2016), 59:1–59:35

    work page 2016

  8. [8]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. 2017. Neural Message Passing for Quantum Chemistry. In 34th International Conference on Machine Learning, ICML 2017 . 1263–1272

    work page 2017

Show all 58 references
  1. [9]

    Hamilton, Zhitao Ying, and Jure Leskovec

    William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive Representa- tion Learning on Large Graphs. In 30th Annual Conference on Neural Information Processing Systems, NeurIPS 2017 . 1024–1034

    2017

  2. [10]

    Kaveh Hassani and Amir Hosein Khas Ahmadi. 2020. Contrastive Multi-View Representation Learning on Graphs. In 37th International Conference on Machine Learning, ICML 2020 . 4116–4126

    2020

  3. [11]

    Yuye He, Sebastien Blandin, Laura Wynter, and Barry Trager. 2014. Analysis and real-time prediction of local incident impact on transportation networks. In 2014 IEEE International Conference on Data Mining Workshop . IEEE, 158–166

    2014

  4. [12]

    Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. 2019. Strategies for pre-training graph neural networks. arXiv preprint arXiv:1905.12265 (2019)

    2019 arXiv

  5. [13]

    Yizhu Jiao, Yun Xiong, Jiawei Zhang, and Yangyong Zhu. 2019. Collective Link Prediction Oriented Network Embedding with Hierarchical Graph Attention. In 28th ACM International Conference on Information and Knowledge Management, CIKM 2019. ACM, 419–428

    2019

  6. [14]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  7. [15]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classi￿cation with Graph Convolutional Networks. In 5th International Conference on Learning Representations, ICLR 2017

    2017

  8. [16]

    Bronstein

    Ron Levie, Federico Monti, Xavier Bresson, and Michael M. Bronstein. 2019. CayleyNets: Graph Convolutional Neural Networks With Complex Rational Spectral Filters. IEEE Trans. Signal Process. 67, 1 (2019), 97–109

    2019

  9. [17]

    Ruoyu Li, Sheng Wang, Feiyun Zhu, and Junzhou Huang. 2018. Adaptive Graph Convolutional Neural Networks. In Thirty-Second AAAI Conference on Arti￿cial Intelligence, AAAI 2018. 3546–3553

    2018

  10. [18]

    Minhua Lin, Zhiwei Zhang, Enyan Dai, Zongyu Wu, Yilong Wang, Xiang Zhang, and Suhang Wang. 2024. Trojan Prompt Attacks on Graph Neural Networks. arXiv preprint arXiv:2410.13974 (2024)

    2024 arXiv

  11. [19]

    Meihan Liu, Zeyu Fang, Zhen Zhang, Ming Gu, Sheng Zhou, Xin Wang, and Jiajun Bu. 2024. Rethinking propagation for unsupervised graph domain adaptation. In Proceedings of the AAAI Conference on Arti￿cial Intelligence, Vol. 38. 13963–13971

    2024

  12. [20]

    Meihan Liu, Zhen Zhang, Jiachen Tang, Jiajun Bu, Bingsheng He, and Sheng Zhou. 2024. Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation. arXiv preprint arXiv:2407.11052 (2024)

    2024 arXiv

  13. [21]

    Shikun Liu, Tianchun Li, Yongbin Feng, Nhan Tran, Han Zhao, Qiang Qiu, and Pan Li. 2023. Structural Re-weighting Improves Graph Domain Adaptation. In International Conference on Machine Learning, ICML 2023 . PMLR, 21778–21793

    2023

  14. [22]

    Shikun Liu, Deyu Zou, Han Zhao, and Pan Li. 2024. Pairwise Alignment Improves Graph Domain Adaptation. arXiv preprint arXiv:2403.01092 (2024)

    2024 arXiv

  15. [23]

    Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. 2015. Learning Transferable Features with Deep Adaptation Networks. In 32nd International Conference on Machine Learning, ICML 2015 . 97–105

    2015

  16. [24]

    Miller McPherson, Lynn Smith-Lovin, and James M Cook. 2001. Birds of a feather: Homophily in social networks. Annual review of sociology 27, 1 (2001), 415–444

    2001

  17. [25]

    Giulia Muzio, Leslie O’Bray, and Karsten Borgwardt. 2021. Biological network analysis with deep learning. Brie￿ngs in bioinformatics 22 (2021), 1515–1530

    2021

  18. [26]

    Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. 2016. Learning Convolutional Neural Networks for Graphs. In 33nd International Conference on Machine Learning, ICML 2016 . 2014–2023

    2016

  19. [27]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al . 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing syste...

    2019

  20. [28]

    Georgios A Pavlopoulos, Maria Secrier, Charalampos N Moschopoulos, Theodoros G Soldatos, Sophia Kossida, Jan Aerts, Reinhard Schneider, and Pan- telis G Bagos. 2011. Using graph theory to analyze biological networks. BioData mining 4 (2011), 1–27

    2011

  21. [29]

    Ziyue Qiao, Xiao Luo, Meng Xiao, Hao Dong, Yuanchun Zhou, and Hui Xiong

  22. [30]

    Xiao Shen, Quanyu Dai, Fu-lai Chung, Wei Lu, and Kup-Sze Choi. 2020. Adversar- ial deep network embedding for cross-network node classi￿cation. In Proceedings of the AAAI conference on arti￿cial intelligence , Vol. 34. 2991–2999

    2020

  23. [31]

    Xiao Shen, Quanyu Dai, Sitong Mao, Fu-lai Chung, and Kup-Sze Choi. 2020. Net- work together: Node classi￿cation via cross-network deep network embedding. IEEE Transactions on Neural Networks and Learning Systems 32 (2020), 1935–1948

    2020

  24. [32]

    Xiran Song, Jianxun Lian, Hong Huang, Zihan Luo, Wei Zhou, Xue Lin, Mingqi Wu, Chaozhuo Li, Xing Xie, and Hai Jin. 2023. xGCN: An Extreme Graph Convo- lutional Network for Large-scale Social Link Prediction. In ACM Web Conference 2023, WWW 2023 . ACM, 349–359

    2023

  25. [33]

    Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, and Zhong Su. 2008. Ar- netMiner: extraction and mining of academic social networks. In 14th ACM International Conference on Knowledge Discovery and Data Mining, SIGKDD 2008 . ACM, 990–998

    2008

  26. [34]

    Eric Tzeng, Judy Ho￿man, Kate Saenko, and Trevor Darrell. 2017. Adversarial Discriminative Domain Adaptation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017 . 2962–2971

    2017

  27. [35]

    Eric Tzeng, Judy Ho￿man, Ning Zhang, Kate Saenko, and Trevor Darrell. 2014. Deep Domain Confusion: Maximizing for Domain Invariance.CoRR abs/1412.3474 (2014)

    2014 arXiv

  28. [36]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In 6th International Conference on Learning Representations, ICLR 2018

    2018

  29. [37]

    Hamilton, Pietro Liò, Yoshua Bengio, and R

    Petar Velickovic, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R. Devon Hjelm. 2019. Deep Graph Infomax. In 7th International Conference on Learning Representations, ICLR 2019

    2019

  30. [38]

    Claudia Wagner, Philipp Singer, Fariba Karimi, Jürgen Pfe￿er, and Markus Strohmaier. 2017. Sampling from Social Networks with Attributes. In 26th Inter- national Conference on World Wide Web, WWW 2017 . 1181–1190

    2017

  31. [39]

    Bingsen Wang, Piero Baraldi, and Enrico Zio. 2023. Deep Multiadversarial Con- ditional Domain Adaptation Networks for Fault Diagnostics of Industrial Equip- ment. IEEE Trans. Ind. Informatics 19, 8 (2023), 8841–8851

    2023

  32. [40]

    Svante Wold, Kim Esbensen, and Paul Geladi. 1987. Principal component analysis. Chemometrics and intelligent laboratory systems 2, 1-3 (1987), 37–52

    1987

  33. [41]

    Ainsworth

    Jun Wu, Jingrui He, and Elizabeth A. Ainsworth. 2023. Non-IID Transfer Learning on Graphs. In Thirty-Seventh AAAI Conference on Arti￿cial Intelligence, AAAI

    2023

  34. [42]

    Man Wu, Shirui Pan, Chuan Zhou, Xiaojun Chang, and Xingquan Zhu. 2020. Unsupervised Domain Adaptive Graph Convolutional Networks. In The Web Conference, WWW 2020 . 1457–1467

    2020

  35. [43]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2019. How Pow- erful are Graph Neural Networks?. In 7th International Conference on Learning Representations, ICLR 2019

    2019

  36. [44]

    Minghao Xu, Hang Wang, and Bingbing Ni. 2022. Graphical modeling for multi- source domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 3 (2022), 1727–1741

    2022

  37. [45]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph Contrastive Learning with Augmentations. In 33rd Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020

    2020

  38. [46]

    Yuning You, Tianlong Chen, Zhangyang Wang, and Yang Shen. 2023. Graph Do- main Adaptation via Theory-Grounded Spectral Regularization. In The Eleventh International Conference on Learning Representations, ICLR 2023

    2023

  39. [47]

    Zafarani and H

    R. Zafarani and H. Liu. 2009. Social Computing Data Repository at ASU. http: //socialcomputing.asu.edu

    2009

  40. [48]

    Muhan Zhang and Yixin Chen. 2018. Link Prediction Based on Graph Neu- ral Networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems, NeurIPS 2018 . 5171–5181

    2018

  41. [49]

    Chawla, and Xiangliang Zhang

    Qiannan Zhang, Shichao Pei, Qiang Yang, Chuxu Zhang, Nitesh V. Chawla, and Xiangliang Zhang. 2023. Cross-Domain Few-Shot Graph Classi￿cation with a Reinforced Task Coordinator. In Thirty-Seventh AAAI Conference on Arti￿cial Intelligence, AAAI 2023. 4893–4901

    2023

  42. [50]

    Xiaowen Zhang, Yuntao Du, Rongbiao Xie, and Chongjun Wang. 2021. Adver- sarial separation network for cross-network node classi￿cation. In Proceedings of the 30th ACM international conference on information & knowledge management . 2618–2626

    2021

  43. [51]

    Jennings

    Youzhi Zhang, Bo An, Long Tran-Thanh, Zhen Wang, Jiarui Gan, and Nicholas R. Jennings. 2017. Optimal Escape Interdiction on Transportation Networks. In WSDM ’25, March 10–14, 2025, Hannover, Germany Yilong Wang, Tianxiang Zhao, Zongyu Wu, and Suhang Wang Twenty-Sixth Internati...

    2017

  44. [52]

    Yanfu Zhang, Shangqian Gao, Jian Pei, and Heng Huang. 2022. Improving Social Network Embedding via New Second-Order Continuous Graph Neural Networks. In KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022 . ...

    2022

  45. [53]

    Yizhou Zhang, Guojie Song, Lun Du, Shuwen Yang, and Yilun Jin. 2019. DANE: Domain Adaptive Network Embedding. In Twenty-Eighth International Joint Conference on Arti￿cial Intelligence, IJCAI 2019 . 4362–4368

    2019

  46. [54]

    Tianxiang Zhao, Dongsheng Luo, Xiang Zhang, and Suhang Wang. 2024. Multi- source Unsupervised Domain Adaptation on Graphs with Transferability Model- ing. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4479–4489

    2024

  47. [55]

    Tianxiang Zhao, Xiang Zhang, and Suhang Wang. 2021. Graphsmote: Imbalanced node classi￿cation on graphs with graph neural networks. In Proceedings of the 14th ACM international conference on web search and data mining . 833–841

    2021

  48. [56]

    Xiaohan Zhao, Bo Zong, Ziyu Guan, Kai Zhang, and Wei Zhao. 2018. Substructure Assembling Network for Graph Classi￿cation. In Thirty-Second AAAI Conference on Arti￿cial Intelligence, AAAI 2018 . 4514–4521

    2018

  49. [57]

    Database

    Chenyi Zhuang and Qiang Ma. 2018. Dual graph convolutional networks for graph-based semi-supervised classi￿cation. In 27th world wide web conference, WWW 2018. 499–508. Ethical Considerations We have thoroughly evaluated the ethical implications and societal impact of our rese...

    2018

  50. [2023]

    arXiv preprint arXiv:2309.10773 (2023)

    Semi-supervised domain adaptation in graph transfer learning. arXiv preprint arXiv:2309.10773 (2023)

    2023 arXiv

Pith tools

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