Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Weak Supervision for Real World 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 A graph neural network that exploits weak-label voting patterns and community structure improves node classification by up to 15 F1 points.

desk verdict Sensible integration of weak supervision and GCL with useful new datasets, but the headline gains are not yet controlled because WSNET's hyperparameters are tuned while baselines run on defaults. read the letter →

arxiv 2506.02451 v1 pith:3GLE2PQG submitted 2025-06-03 cs.LG

Pratheeksha Nair , Reihaneh Rabbany This is my paper
classification cs.LG
keywords weaksupervisiongraphcontrastivelearningnodeclassificationnoisylabelslabelingfunctionshumantraffickingdetectionmisinformationneuralnetworks
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 claims that noisy, indirect supervision signals—weak labels from multiple heuristics or expert rules—can be turned into accurate node classifiers on real-world graphs if they are combined with graph contrastive learning. It presents WSNET, a graph neural network trained with three objectives: a weighted cross-entropy on majority-vote weak labels, an InfoNCE contrastive loss that treats nodes with similar weak-label voting patterns as positives, and a structure-based contrastive loss that aligns each node with its community embedding. Across four real-world datasets (human trafficking ads, misinformation statements, citation networks) and five synthetic benchmarks with controlled label noise, WSNET reports weighted F1 gains of up to 15 points over contrastive and noisy-label baselines. The authors position this as a step toward using weak supervision where high-quality labels are expensive or slow to obtain, such as trafficking detection and fact-checking.

What carries the argument

The load-bearing object is the three-term loss $L = L_{\text{SCon}} + L_{\text{WLCE}} + L_{\text{WLCon}}$. $L_{\text{WLCE}}$ is a cross-entropy on majority-vote labels weighted by node reliability (weak-label agreement and closeness to the node's embedding cluster centroid); $L_{\text{WLCon}}$ is an InfoNCE loss in which each node's positive pair is sampled by cosine similarity of its weak-label vector, on the evidence that similar voting patterns indicate same-class membership; $L_{\text{SCon}}$ contrasts each node embedding with its community's mean-pooled embedding against corrupted views, injecting graph structure as a regularizer. The combination is what aligns representations with classes despite imperfect supervision.

What would settle it

Train WSNET on a controlled graph where labeling functions are engineered so that voting-pattern similarity is independent of true class (e.g., all LFs are random with identical error rates and shared biases), and compare it to a GNN trained on the same weak labels without the contrastive losses; if full WSNET does not beat that GNN, the weak-label contrastive assumption is not doing the work the paper claims.

Watch

Extended reading notes

Core claim

On its own terms, WSNET's central discovery is that weak-label voting patterns, despite being noisy, carry reliable class information that contrastive learning can exploit: nodes whose weak-label distributions are similar belong to the same class more often than not, so pulling such nodes together in embedding space improves classification. The paper shows this empirically (its Figure 3) and uses it to define a weak-label contrastive loss. To keep noisy labels from distorting the representation, WSNET adds a structure-based contrastive loss that contrasts each node with its graph community, acting as a denoising regularizer, and a weighted cross-entropy loss that down-weights unreliable nodes determined by weak-label entropy and embedding representativeness. The integrated loss lets WSNET outperform not only self-supervised graph methods and label models but also noisy-label learning methods that try to clean labels, with the largest gains at low weak-label accuracy.

Load-bearing premise

The method assumes nodes with similar weak-label voting patterns are usually the same true class; if noisy labeling functions produce correlated voting patterns across different classes, the weak-label contrastive loss will pull distinct classes together and hurt classification.

Editorial extensions

If this is right

  • WSNET reports consistent gains over contrastive, noisy-label, and weak-supervision baselines, up to 15 points in weighted F1, with the largest advantage when weak labels are least accurate (label accuracy as low as 0.1).
  • On LIAR-WS, WSNET reaches 88% F1, above the best single weak label's 87% accuracy, indicating the framework denoises rather than merely echoes its supervision.
  • The method scales to large graphs: it runs on ogbn-arxiv (169k nodes) where several noisy-label and supervised-contrastive baselines run out of memory.
  • Two new weakly supervised graph benchmarks are introduced, LIAR-WS for misinformation and CORA-WS for citation topic classification, usable by the community.
  • Ablations show removing the weak-label cross-entropy term causes the largest drop, while each contrastive component contributes smaller but consistent gains.

Reading between the lines

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

  • If weak-label similarity reliably tracks class membership, the same positive-sampling idea could transfer to other data types with multiple noisy annotators, where a vector of annotator votes replaces the labeling-function matrix.
  • The method's success depends on labeling functions not sharing systematic biases; constructing LFs with a common spurious keyword would create high weak-label similarity across different true classes and could serve as a stress test of the weak-label contrastive assumption.
  • The structure-based regularizer suggests a testable prediction: on graphs without meaningful community structure, its contribution should shrink, consistent with the paper's own limitation note about irregular biological graphs.
  • Because hyperparameters tau near 0.5 and r near 50 work across datasets, a theory for why moderate negative-sample counts balance contrastive stringency with stability under weak labels could be explored.
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 WSNET, a graph neural network framework for weakly supervised node classification. The model combines three loss terms: a weighted cross-entropy loss on majority-vote weak labels, an InfoNCE contrastive loss that pulls together nodes with similar weak-label voting patterns, and a structure-based contrastive loss that contrasts node embeddings with community representations. The authors evaluate WSNET on four real-world graphs (LIAR-WS, ASW-REAL, ASW-SYNTH, CORA-WS) and five synthetic-noise benchmark datasets, reporting consistent F1 improvements over SSL, noisy-label learning, programmatic weak supervision, and supervised contrastive baselines, with claims of up to 15% absolute F1 gains. An ablation study decomposes the contribution of each loss term, and the paper releases code and two new datasets.

Significance. If the comparative results are reliable, WSNET is a relevant contribution at the intersection of programmatic weak supervision and graph contrastive learning. The method is simple and reproducible, the public code release is a strength, and the newly introduced LIAR-WS and CORA-WS datasets may be useful to the community. However, the central quantitative claim—consistent superiority over state-of-the-art methods by up to 15% F1—rests on an asymmetric evaluation protocol in which WSNET's hyperparameters are tuned on validation while all baselines use defaults. Until this is controlled, the headline advantage is not established. The ablation study is informative but lacks a plain weak-label-trained GCN baseline, so the marginal contribution of the contrastive components is not isolated.

major comments (3)
  1. [§5.1, Table 2, Figure 5] The comparison is not controlled: WSNET's temperature tau and negative-sample count r are fine-tuned on a validation set, while all baselines are run with official implementations and default hyperparameters. Contrastive methods such as SupCon, ClusterSCL, GRACE, and GCA are known to be sensitive to tau and r, so the reported 'up to 15%' advantage may be an artifact of tuning asymmetry rather than a property of the method. Please retune baselines under the same validation procedure, or justify the default choices with sensitivity analyses, before claiming consistent state-of-the-art performance.
  2. [§4, Eq. (2), Table 3] The ablation study shows that removing the weak-label cross-entropy term causes the largest performance drop (e.g., LIAR-WS from 0.88 to 0.54 and CORA-WS from 0.40 to 0.15), yet no baseline of a plain GCN trained only on weighted majority-vote weak labels is included in Table 2 or Figure 5. Without this control, the reader cannot tell whether the gain of WSNET over baselines comes from the weighted CE term alone or from the contrastive components. Please add such a baseline and report its performance in the main tables.
  3. [§5.2, Table 1, Table 3] On ASW-REAL, ASW-SYNTH, and CORA-WS, the mean accuracy of the weak labels is near random (11.0%, 10.1%, and 11.1%, respectively), yet WSNET reports large F1 gains (0.72, 0.78, and 0.40). The paper does not explain where this signal comes from, and the ablation table omits these datasets, so the mechanism behind the gains is unclear. Please provide per-dataset ablations and an analysis separating the contribution of graph structure from that of the weak labels; without this, the claim that weak supervision drives performance on these datasets is not supported.
minor comments (6)
  1. [§4, Eq. (3)] The notation for negative samples is inconsistent: the text defines S^-_j as a set of r samples, but the denominator uses exp(h_i · S^-_j / tau), which suggests a vector. Please denote the j-th negative sample as s^-_j and write exp(h_i · s^-_j / tau).
  2. [§6.1, Table 2] The text states that WSNET achieves 78% and 74% on ASW-SYNTH and ASW-REAL, but Table 2 lists 0.78 ± 0.08 and 0.72 ± 0.03; please reconcile the discrepancy.
  3. [§6.2] The paragraph refers to 'WS-GCL' when describing WSNET's advantages; this appears to be a typo.
  4. [Table 1] The benchmark dataset rows do not align with the column headers (Mean Acc., Max Acc., Cov., m); for Citeseer, the values '10-100 10-100 70.0 10' are ambiguous. Please clarify what is reported for synthetic weak labels.
  5. [Throughout] The method name is typeset inconsistently as 'WSN ET' in most of the text and 'WSNET' in the abstract and tables; please standardize the spelling.
  6. [§7] The limitations paragraph does not address the potential failure mode of the weak-label contrastive loss when nodes from different classes have similar weak-label voting patterns. Given that the method assumes such similarity indicates same-class membership, a brief discussion or empirical check would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: WSNET's reported gains are evaluated on held-out ground truth, and the weak-label and contrastive losses are training objectives rather than derived predictions.

full rationale

WSNET's loss (Eq. 5) combines weak-label cross-entropy (Eq. 2), weak-label contrastive InfoNCE (Eq. 3), and structure-based contrastive (Eq. 4), all computed from the weak-label matrix, node features, and graph adjacency. The central quantitative claim is test weighted F1 on held-out ground-truth labels under 80-10-10 splits (Section 5), which is external to every loss term; no reported number is obtained by re-substituting the fitted weak labels or validation-tuned hyperparameters into the test metric. The weak-label contrastive component does select positives from the same weak-label signal used by the classification loss, so it is not independent evidence of class membership, but the paper uses it only as a training heuristic and does not claim to derive ground truth from it; this is a robustness concern, not a circular derivation. The self-citation [5] is used to corroborate the entropy-reliability observation and as a domain-specific comparison, not as the sole justification for the method. The hyperparameter tuning asymmetry (WSNET's r and tau tuned on validation while baselines use defaults) is an experimental fairness issue, not circularity. Therefore the derivation chain is self-contained with respect to its evaluation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method relies on two empirical heuristics (entropy-reliability and weak-label-similarity as class-similarity), a community-structure assumption, and the use of majority-vote labels as training targets. These are domain assumptions with limited theoretical backing.

free parameters (2)
  • temperature tau = ~0.5 (typical, per-dataset)
    Temperature in InfoNCE loss, fine-tuned on validation set; Figure 4 shows optimal near 0.5.
  • number of negative samples r = ~50 (typical, per-dataset)
    Negative sample count in InfoNCE, fine-tuned on validation set; Figure 4 shows optimal near 50.
assumptions (4)
  • domain assumption Low weak-label entropy implies higher label reliability
    Used to weight cross-entropy loss in Eq. 1; supported by Figure 2 but not guaranteed across applications.
  • domain assumption Similar weak-label distributions imply same class
    Used to sample positive pairs for InfoNCE in Eq. 3; supported by Figure 3 but is an empirical heuristic.
  • domain assumption Graph communities capture class-relevant structure
    Structure-based contrastive loss in Eq. 4 pools over communities; assumes community structure aligns with classes.
  • domain assumption Majority vote is a sufficient label aggregation for training
    Method uses majority vote of weak labels as classification targets; no label model is used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weak Supervision for Real World Graphs." pith.science (2026). https://pith.science/paper/3GLE2PQG

@misc{pith2026250602451,
  author       = {Pith},
  title        = {Pith review of: Weak Supervision for Real World Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GLE2PQG}},
  note         = {Machine review of arXiv:2506.02451}
}
read the original abstract

Node classification in real world graphs often suffers from label scarcity and noise, especially in high stakes domains like human trafficking detection and misinformation monitoring. While direct supervision is limited, such graphs frequently contain weak signals, noisy or indirect cues, that can still inform learning. We propose WSNET, a novel weakly supervised graph contrastive learning framework that leverages these weak signals to guide robust representation learning. WSNET integrates graph structure, node features, and multiple noisy supervision sources through a contrastive objective tailored for weakly labeled data. Across three real world datasets and synthetic benchmarks with controlled noise, WSNET consistently outperforms state of the art contrastive and noisy label learning methods by up to 15% in F1 score. Our results highlight the effectiveness of contrastive learning under weak supervision and the promise of exploiting imperfect labels in graph based settings.

Figures

Figures reproduced from arXiv: 2506.02451 by the authors.

Figure 1
Figure 1. The graph (A, X) and weak labels (Λ) are input to the WSNET pipeline which then produces label predictions. Supervised Graph Contrastive Learning: SupCon [13] is a supervised contrastive learning method for ImageNet classification and was adapted to graphs in ClusterSCL [14]. It uses class label information in the contrastive loss to learn efficient embeddings. To negate the impacts of SupCon induced by the intra-cl… view at source ↗
Figure 2
Figure 2. Nodes whose majority vote aggregated label corresponds to the ground truth (‘Correct agg.’) have a higher [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Node pairs with highly similar weak label distributions (chosen as positives) belong to the same class more often than pairs with low weak label similarity (negatives). are computed using Equation 1. ρi = |Qi | ∗ (hi · hQi ) PN j=1(hj · hQj ) ∗ entropy(Λi) (1) where hi is the hidden representation of node i and hQi is the centroid representation of cluster Qi that i belongs to. ‘.’ indicates cosine similarity. The w… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: We fine-tuned tau and r on a validation set to choose the optimal hyperparameter values for each dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Performance of WSNET on benchmark graph datasets. The y-axis is the weighted F1 score and the x-axis is the label accuracy. WSNET has a clear advantage compared to baselines particularly when label accuracy is low. SupCon and PI-GNN resulted in OOM for Coauthor and HLM…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages

  1. [1]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017

    2017

  2. [2]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In ICLR, 2018

    work page 2018

  3. [3]

    Multi-stage self-supervised learning for graph convolutional networks on graphs with few labeled nodes

    Ke Sun, Zhouchen Lin, and Zhanxing Zhu. Multi-stage self-supervised learning for graph convolutional networks on graphs with few labeled nodes. In AAAI, volume 34, 2020

    work page 2020

  4. [4]

    Effective Stabilized Self-Training on Few-Labeled Graph Data

    Ziang Zhou, Jieming Shi, Shengzhong Zhang, Zengfeng Huang, and Qing Li. Effective semi-supervised node classification on few-labeled graph data. arXiv:1910.02684, 2019

    work page Pith review arXiv 1910

  5. [5]

    T-net: Weakly supervised graph learning for combatting human trafficking

    Pratheeksha Nair, Javin Liu, Catalina Vajiac, Andreas Olligschlaeger, et al. T-net: Weakly supervised graph learning for combatting human trafficking. In AAAI, volume 38, 2024

    work page 2024

  6. [6]

    Scaling up fact-checking using the wisdom of crowds

    Jennifer Allen, Antonio A Arechar, Gordon Pennycook, and David G Rand. Scaling up fact-checking using the wisdom of crowds. Science advances, 7(36):eabf4393, 2021. 9 A PREPRINT - SEPTEMBER 3, 2025

    work page 2021

  7. [7]

    Justice in misinformation detection systems: An analysis of algorithms, stakeholders, and potential harms

    Terrence Neumann, Maria De-Arteaga, and Sina Fazelpour. Justice in misinformation detection systems: An analysis of algorithms, stakeholders, and potential harms. FAccT ’22, 2022

    work page 2022

  8. [8]

    Data programming: Creating large training sets, quickly

    Alexander J Ratner, Christopher M De Sa, Sen Wu, Daniel Selsam, and Christopher Ré. Data programming: Creating large training sets, quickly. NeurIPS, 29, 2016

    work page 2016

Show all 45 references
  1. [9]

    Learning hyper label model for programmatic weak supervision

    Renzhi Wu, Shen-En Chen, Jieyu Zhang, and Xu Chu. Learning hyper label model for programmatic weak supervision. In ICLR, 2023

    2023

  2. [10]

    Bigbio: a framework for data-centric biomedical natural language processing

    Jason Fries, Leon Weber, Natasha Seelam, et al. Bigbio: a framework for data-centric biomedical natural language processing. NeurIPS, 35, 2022

    2022

  3. [11]

    Resonant anomaly detection with multiple reference datasets

    Mayee F Chen, Benjamin Nachman, and Frederic Sala. Resonant anomaly detection with multiple reference datasets. Journal of High Energy Physics , 2023(7), 2023

    2023

  4. [12]

    Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm

    Petar Veliˇckovi´c, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. Deep graph infomax. In ICLR, 2019

    2019

  5. [13]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. arXiv:2004.11362, 2020

    2004 arXiv

  6. [14]

    Clusterscl: cluster-aware supervised contrastive learning on graphs

    Yanling Wang, Jing Zhang, Haoyang Li, et al. Clusterscl: cluster-aware supervised contrastive learning on graphs. In Proceedings of the ACM Web Conference 2022, 2022

    2022

  7. [15]

    Automating the construction of internet portals with machine learning

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automating the construction of internet portals with machine learning. Information Retrieval, 3, 2000

    2000

  8. [16]

    liar, liar pants on fire

    William Yang Wang. “liar, liar pants on fire”: A new benchmark dataset for fake news detection. In Regina Barzilay and Min-Yen Kan, editors, Proceedings of the 55th Annual Meeting of the Association for Computa- tional Linguistics (V olume 2: Short Papers) , pages 422–426, Van...

    2017

  9. [17]

    A survey on programmatic weak supervision

    Jieyu Zhang, Cheng-Yu Hsieh, Yue Yu, Chao Zhang, and Alexander Ratner. A survey on programmatic weak supervision. arXiv:2202.05433, 2022

    2022 arXiv

  10. [18]

    Training complex models with multi-task weak supervision

    Alexander Ratner, Braden Hancock, Jared Dunnmon, Frederic Sala, Shreyash Pandey, and Christopher Ré. Training complex models with multi-task weak supervision. In AAAI, volume 33, 2019

    2019

  11. [19]

    Multi-resolution weak supervision for sequential data

    Paroma Varma, Frederic Sala, Shiori Sagawa, Jason Fries, Daniel Fu, Saelig Khattar, Ashwini Ramamoorthy, Ke Xiao, Kayvon Fatahalian, James Priest, et al. Multi-resolution weak supervision for sequential data. NeurIPS, 32, 2019

    2019

  12. [20]

    Fast and three-rious: Speeding up weak supervision with triplet methods

    Daniel Fu, Mayee Chen, Frederic Sala, Sarah Hooper, Kayvon Fatahalian, and Christopher Ré. Fast and three-rious: Speeding up weak supervision with triplet methods. In ICML. PMLR, 2020

    2020

  13. [21]

    Nrgnn: Learning a label noise-resistant graph neural network on sparsely and noisily labeled graphs

    Enyan Dai, Charu Aggarwal, and Suhang Wang. Nrgnn: Learning a label noise-resistant graph neural network on sparsely and noisily labeled graphs. arXiv:2106.04714, 2021

    2021 arXiv

  14. [22]

    Noise-robust graph learning by estimating and leveraging pairwise interactions

    Xuefeng Du, Tian Bian, Yu Rong, Bo Han, et al. Noise-robust graph learning by estimating and leveraging pairwise interactions. TMLR, 2023

    2023

  15. [23]

    Learning on graphs under label noise

    Jingyang Yuan, Xiao Luo, Yifang Qin, Yusheng Zhao, Wei Ju, and Ming Zhang. Learning on graphs under label noise. In IEEE ICASSP, 2023

    2023

  16. [24]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning . PMLR, 2020

    2020

  17. [25]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. NeurIPS, 33, 2020

    2020

  18. [26]

    Contrastive multi-view representation learning on graphs

    Kaveh Hassani and Amir Hosein Khasahmadi. Contrastive multi-view representation learning on graphs. In International conference on machine learning . PMLR, 2020

    2020

  19. [27]

    Deep graph contrastive representation learning

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Deep graph contrastive representation learning. ICML Graph Representation Learning and Beyond , 2020

    2020

  20. [28]

    Csgcl: Community-strength- enhanced graph contrastive learning

    Chen Han, Zhao Ziwen, Li Yuhua, Zou Yixiong, Li Ruixuan, and Zhang Rui. Csgcl: Community-strength- enhanced graph contrastive learning. CoRR, abs/2305.04658, 2023

    2023 arXiv

  21. [29]

    Graph contrastive learning with adaptive augmentation

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Graph contrastive learning with adaptive augmentation. In Proceedings of the Web Conference 2021, 2021

    2021

  22. [30]

    Large-scale representation learning on graphs via bootstrapping

    Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, et al. Large-scale representation learning on graphs via bootstrapping. In ICLR, 2022. 10 A PREPRINT - SEPTEMBER 3, 2025

    2022

  23. [31]

    Graph Representation Learning via Graphical Mutual Information Maximization

    Zhen Peng, Wenbing Huang, Minnan Luo, Qinghua Zheng, Yu Rong, Tingyang Xu, and Junzhou Huang. Graph Representation Learning via Graphical Mutual Information Maximization. In Proceedings of The Web Conference, 2020

    2020

  24. [32]

    Simple unsupervised graph representation learning

    Yujie Mo, Liang Peng, Jie Xu, Xiaoshuang Shi, and Xiaofeng Zhu. Simple unsupervised graph representation learning. In AAAI, 2022

    2022

  25. [33]

    Augmentation-free graph contrastive learning of invariant-discriminative representations.IEEE Transactions on Neural Networks and Learning Systems, 2023

    Haifeng Li, Jun Cao, Jiawei Zhu*, Qinyao Luo, Silu He, and Xuying Wang. Augmentation-free graph contrastive learning of invariant-discriminative representations.IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  26. [34]

    Kefato and Sarunas Girdzijauskas

    Zekarias T. Kefato and Sarunas Girdzijauskas. Self-supervised graph neural networks without explicit negative sampling, 2021

    2021

  27. [35]

    Jgcl: Joint self-supervised and supervised graph contrastive learning

    Selahattin Akkas and Ariful Azad. Jgcl: Joint self-supervised and supervised graph contrastive learning. In Companion Proceedings of the Web Conference 2022 , 2022

    2022

  28. [36]

    Weakly supervised contrastive learning

    Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021

    2021

  29. [37]

    Rethinking weak supervision in helping contrastive learning

    Jingyi Cui, Weiran Huang, Yifei Wang, and Yisen Wang. Rethinking weak supervision in helping contrastive learning. arXiv:2306.04160, 2023

    2023 arXiv

  30. [38]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018

    2018 arXiv

  31. [39]

    Combating misinformation in the age of llms: Opportunities and challenges

    Canyu Chen and Kai Shu. Combating misinformation in the age of llms: Opportunities and challenges. AI Magazine, 2023

    2023

  32. [40]

    Towards reliable misinformation mitigation: Generalization, uncertainty, and gpt-4

    Kellin Pelrine, Anne Imouza, Camille Thibault, Meilina Reksoprodjo, Caleb Gupta, Joel Christoph, Jean-François Godbout, and Reihaneh Rabbany. Towards reliable misinformation mitigation: Generalization, uncertainty, and gpt-4. arXiv:2305.14928, 2023

    2023 arXiv

  33. [41]

    Citeseer: An automatic citation indexing system

    C Lee Giles, Kurt D Bollacker, and Steve Lawrence. Citeseer: An automatic citation indexing system. In Proceedings of the third ACM conference on Digital libraries , 1998

    1998

  34. [42]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. NeurIPS, 32, 2019

    2019

  35. [43]

    Microsoft academic graph: When experts are not enough

    Kuansan Wang, Zhihong Shen, Chiyuan Huang, Chieh-Han Wu, Yuxiao Dong, and Anshul Kanakia. Microsoft academic graph: When experts are not enough. Quantitative Science Studies, 2020

    2020

  36. [44]

    Image-based recommendations on styles and substitutes

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In ACM SIGIR, 2015

    2015

  37. [45]

    Graphmix: Improved training of gnns for semi-supervised learning

    Vikas Verma, Meng Qu, Kenji Kawaguchi, Alex Lamb, Yoshua Bengio, et al. Graphmix: Improved training of gnns for semi-supervised learning. In AAAI, volume 35, 2021. 11

    2021

Pith tools

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