CoATA: Effective Co-Augmentation of Topology and Attribute for Graph Neural Networks

Longlong Lin; Tao Jia; Tao Liu; Xi Ou; Youan Zhang; Yunfeng Yu; Zhiqiu Ye

REVIEW 5 major objections 6 minor 64 references

CoATA: Effective Co-Augmentation of Topology and Attribute for Graph Neural Networks

T0 review · 5 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 CoATA co-augments topology and attributes to outperform single-dimension graph augmentation on six of seven graphs.

desk verdict A promising co-augmentation pipeline undercut by an algorithm that doesn't compute the PPR it claims. read the letter →

arxiv 2506.22299 v1 pith:QZDPQJBE submitted 2025-06-27 cs.LG cs.AI

Tao Liu , Longlong Lin , Yunfeng Yu , Xi Ou , Youan Zhang , Zhiqiu Ye , Tao Jia This is my paper

classification cs.LGcs.AI

keywords graphneuralnetworksaugmentationtopologyandattributeco-augmentationPersonalizedPageRanknode-attributebipartiteheterophilysemi-supervisednodeclassificationprototypealignment

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

CoATA sets out to establish that graph augmentation is more effective when topology and node attributes are corrected together rather than in isolation. The pipeline first propagates structural signals into node features with a residual mechanism, then rebuilds or refines edges by running Personalized PageRank on a node-attribute bipartite graph built from those enriched features, and finally trains a shared-weight GNN on the original and co-augmented graphs with a prototype-alignment loss. On seven benchmark graphs, the paper reports that this approach outperforms eleven comparison methods on six datasets, with accuracy improvements between 1.1% and 17.9%. The significance is that it targets the dilemma loop in which missing edges block feature propagation and noisy features mislead topology repair.

What carries the argument

The central object is a node-attribute bipartite graph with an intra-set push-based Personalized PageRank (PPR) procedure. In this bipartite graph, one side is the node set and the other side is the feature-dimension set, with edge weights taken from the topology-enriched attribute matrix; the intra-set push algorithm propagates probability mass along node-to-feature-to-node walks and returns PPR scores only on the node side. Those scores form a co-augmented adjacency matrix, built either by keeping top-K neighbors per node or by adding and removing a fixed number of edges. The supporting machinery is the topology-enriched attribute module, a residual multi-hop propagation with mixing coefficient $\beta$, and the prototype-alignment contrastive loss that keeps the original and augmented channels aligned at the class level.

What would settle it

One settling test: on a synthetic graph with known latent classes and injected feature noise, compute whether the edges CoATA adds via bipartite PPR connect nodes of the same latent class at a rate exceeding chance; if the added edges are no better than random rewiring, the co-augmentation mechanism is not carrying the result.

Watch

Extended reading notes

Core claim

The central discovery is that topology and attribute augmentation can be chained into a mutual-correction loop, and that this loop is what carries the accuracy gains. Structure-enriched attributes give the edge-reconstruction step a cleaner signal than raw features, while the reconstructed edges give distant but semantically related nodes a path to exchange information; the dual-channel setup keeps the original edges as a faithful local view. The load-bearing comparison is against single-dimensional augmenters: the paper reports that omitting either the topology-enriched attributes or the attribute-informed topology step degrades performance, and that the full CoATA is the top performer on the heterophilic graphs Squirrel and Chameleon. The authors' stated conclusion is that co-augmentation captures the synergy between topology and attributes that single-dimension methods miss.

Load-bearing premise

The load-bearing premise is that Personalized PageRank over a node-attribute bipartite graph, built from topology-enriched features, produces pairwise similarities that reflect true semantic relatedness better than the original adjacency or raw-feature cosine similarity.

Editorial extensions

If this is right

On the five homophilous datasets, CoATA reports the highest accuracy on four, with a 0.56-point or larger margin over the nearest competitor on Coauthor-Phy.
On the two heterophilic graphs, Squirrel and Chameleon, CoATA outperforms every baseline by at least 1.31 and 1.40 percentage points, respectively, which the authors attribute to the residual-propagation step countering misleading local neighborhoods.
The ablation table shows that adding the topology-enriched attribute step to the structural-only module raises Citeseer accuracy from 71.07% to 74.90%, and the full pipeline reaches 75.37%; removing either the consistency loss or the augmented-channel supervision drops accuracy noticeably.
Because the augmented adjacency can be constructed by fixed-size KNN or by adding and removing a fixed number of edges per node, the framework offers two concrete graph-reconstruction strategies that share the same training procedure.
The results imply that a GNN trained on both the raw and the co-augmented graph inherits robustness from each view, since the two channels share parameters.

Reading between the lines

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

An unstated extension is to treat the bipartite-PPR scores as edge logits for supervised link prediction, rather than as a fixed preprocessing step; the co-augmented adjacency is already a dense node-proximity matrix that could be thresholded differently per task.
Because the channels share one encoder, any future improvement to the base GNN would likely transfer directly to CoATA, so the augmentation stage could be tested on top of stronger encoders than the GCN used here.
The fixed-size graph reconstruction strategies (top-K and add/remove) assume an approximately uniform connectivity budget per node; a natural next step would make the number of added edges adaptive or differentiable, which the paper leaves implicit.
The reported PubMed shortfall hints that low-dimensional features dampen the benefit of topology-enriched attributes; a direct test would be to run CoATA on a low-dimensional version of a high-dimensional dataset and compare the gap.

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.

Desk Editor's Note

read the letter

Two things you should know about CoATA. First, the co-augmentation idea is genuinely interesting: propagate structural info to denoise attributes, then use those enriched attributes to rebuild the graph via a proximity measure on a node-attribute bipartite graph, then align prototypes across the two channels. The gains on heterophilic benchmarks (Squirrel +1.3, Chameleon +1.4 over GeoMix) are real and worth taking seriously. Second, the paper's central mechanism is misdescribed. Algorithm 2 is claimed to compute Personalized PageRank on the bipartite graph, but it drops the alpha-teleport at attribute nodes. On a tiny graph with V={s,v1}, U={a}, edges s-a and a-v1, alpha=0.5, exact PPR gives pi_s=7/12, pi_v1=1/12; Algorithm 2 converges to pi_s=0.75, pi_v1=0.25. This isn't a rounding issue; the algorithm is solving a different recurrence. Consequently Theorem 4.2 (whose proof is in an unavailable supplementary) cannot be applied to its output, and the paper's narrative that AIT uses 'the celebrated proximity metric Personalized PageRank' is inaccurate.

The empirical evaluation is solid in the sense that it compares against eleven baselines on seven datasets, but the abstract's 'outperforms eleven state-of-the-art baseline methods' is contradicted by the paper's own PubMed row (80.21 vs 80.73 S3-CL and 80.84 SimP-GCN). The ablation text also misattributes the +4.39% gain on Citeseer to DPA when that's the cumulative improvement over GCN, not the incremental gain. The theorems themselves are weak: 4.1 is imported verbatim from GeoMix, and 4.3 restates the contrastive loss definition. All proofs are deferred to a supplementary that isn't available.

Proportionately, though, the soft spots are fixable. If Algorithm 2 were corrected to actually implement PPR (by adding a teleport at U), or if the paper were honest that it computes a related V-only proximity measure and re-derived the guarantee, the empirical contribution would stand. As written, the central algorithm does not match its formal description.

This paper deserves a serious referee, not a desk reject. A referee would insist on the algorithm correction, the proofs, and a toned-down abstract. For my own work, I wouldn't cite it as a PPR-based method until the discrepancy is resolved; the co-augmentation idea is promising but needs a cleaner write-up. I'd bring it to the reading group as a case study in how pseudocode can drift from its claimed invariant.

Referee Report

5 major / 6 minor

Summary. The paper proposes CoATA, a dual-channel GNN for node classification that jointly augments node attributes and graph topology. The TEA module propagates features over the original graph with a residual term, the AIT module constructs a node-attribute bipartite graph from the enriched features and uses a push-based algorithm that the paper calls Personalized PageRank to build a co-augmented adjacency matrix, and the DPA module trains a shared GNN on the original and augmented views with cross-entropy, consistency, and prototype-contrastive losses. Experiments on seven benchmark datasets against eleven baselines report state-of-the-art or competitive accuracy, with particularly strong gains on the two heterophilic graphs.

Significance. The high-level idea of co-augmenting features and topology in a closed loop, and the dual-channel training with prototype alignment, is timely and the reported gains on Squirrel and Chameleon are interesting. The paper's stated theoretical anchors, however, are not currently reliable: the AIT algorithm is not the PPR procedure described in the text, and the proofs of Theorems 4.2 and 4.3 are either absent or circular. The empirical evaluation is broad (seven datasets, eleven baselines, ablations), but the headline claim in the abstract is contradicted by the paper's own PubMed results, and the ablation discussion misreports which module produces the largest gain. If the algorithmic description and theoretical statements are corrected, the proposed framework has potential; in its present form the central mechanism is misdescribed.

major comments (5)

[Section 4.2, Theorem 4.2] Algorithm 2 does not compute the Personalized PageRank defined by Eq. (6) on the node-attribute bipartite graph G_b. In standard PPR, every push of a node v adds α·r to π(v) and distributes (1−α)·r to its neighbors; Algorithm 2 applies the α teleport only when pushing V nodes, while residue flushed from U nodes is sent entirely back to V neighbors with no teleport and no possibility of stopping at U. Concretely, for V={s,v1}, U={a}, unit-weight edges s–a and a–v1, and α=0.5, the exact PPR from Eq. (6) gives π_s=7/12 and π_v1=1/12 (before normalization over V), whereas Algorithm 2 converges to π_s=3/4 and π_v1=1/4, a ratio of 3 instead of 7. The matrix returned by Algorithm 2 is the PPR of a different walk, namely the two-step V-projection with transition T_{v→v'} = (1−α)·Σ_u (w(v,u)/d(v))·(w(v',u)/d(u)) and teleport to the source at V. Since AIT is described throughout the paper as 'PPR on the node-attribute bipartite graph' (Sections 1, 4.2, and 6), the method as written is not the method that the theoretical discussion and Theorem 4.2 describe. Please either implement true bipartite PPR and restrict to V, or explicitly redefine AIT as PPR on the V-projection and adapt the theory accordingly.
[Section 4.2, Theorem 4.2] The proof of Theorem 4.2 is deferred to a Supplementary Material that is not included with the submitted manuscript, and the statement is about PPR on the bipartite graph G_b. Since Algorithm 2 does not compute that PPR (see the previous comment), even a correct proof of Theorem 4.2 would not bound the scores that are actually used to reconstruct the adjacency matrix. Please provide a proof that applies to the implemented walk, or replace the theorem with a statement and proof for the actual two-step walk.
[Abstract and Table 2] The abstract states that CoATA 'outperforms eleven state-of-the-art baseline methods', but Table 2 shows that on PubMed CoATA (80.21±0.16) trails S3-CL (80.73±0.28) and SimP-GCN (80.84±0.17). The supported claim is that CoATA achieves the best accuracy on six of the seven datasets; please revise the abstract and Section 1 accordingly.
[Section 5.3, Table 3] The ablation paragraph misattributes the gains. On Citeseer, adding DPA (O3) to O1+O2 increases accuracy by only 75.37−74.90 = 0.47 percentage points, whereas the text reports '+4.39%' for this step; that value is the full-model gain over GCN. Similarly, the 'accuracy improvement ... on Citeseer (+3.92%)' attributed to TEA is the O1+O2 gain over GCN rather than the incremental gain over O1, which is 74.90−71.07 = 3.83 points. The conclusion that DPA provides the largest performance boost is not supported by the table; please report incremental gains separately.
[Section 5.1 and reproducibility] The paper references a Supplementary Material for dataset details, implementation details, and proofs, but this material is not present in the submitted manuscript. The hyperparameter section gives only ranges (α∈[0.1,0.9], β∈[0.1,0.9], h∈[1,4]) and omits the values of K, k_add, k_del, λ1, λ2, λ3, and per-dataset choices. Without these details or a reference to a complete supplement, the experimental results are not reproducible.

minor comments (6)

[Introduction] The phrase 'natural language processing natural language processing' contains a duplicated fragment; remove the repetition.
[Figure 5] The axis labels in Figure 5 contain garbled Unicode sequences (e.g., '/uni00000024/...'), making the plots illegible; please regenerate the figure.
[Section 4.3, Theorem 4.3] Theorem 4.3 is essentially a restatement of the definition of L_dpa rather than a substantive guarantee; recast it as a remark or provide a real proof.
[Table 1] Table 1 spells the dataset as 'Chamelon' while the text uses 'Chameleon'; make the spelling consistent.
[References] Reference [26] is a duplicate of [25] (both are the APPNP paper); merge the citations.
[Section 2 and Table 2] Baseline names are inconsistent across the paper (e.g., 'PSAGNN' vs 'PSA-GNN' and 'Simp-GCN' vs 'SimP-GCN'); use a single name per method throughout.

Circularity Check

1 steps flagged · score 1.0 of 10

One minor self-definitional theorem in the theoretical packaging; the empirical pipeline is self-contained and not circular.

self definitional [Section 4.3, Theorem 4.3 (after Eq. 11)]
"Theorem 4.3 (Prototype Alignment Guarantee). Minimizing the contrastive loss L_dpa ensures that prototypes from the same class across channels are pulled closer and those from different classes remain separated."

The loss L_dpa is defined in Eq. (11) exactly as a negative log-ratio of same-class prototype similarity f(p_j, p'_j) to cross-class similarities f(p_j, p'_q) and f(p_q, p'_j). Minimizing this loss by construction increases the same-class numerator and decreases the cross-class denominator, which is precisely the statement that same-class prototypes are pulled closer and different-class prototypes are separated. The theorem therefore restates the definition of the contrastive objective rather than deriving an independent guarantee, and its proof is deferred to the supplementary material. This is a self-definitional packaging step, but it is not load-bearing for the empirical node-classification results, which stand on the reported experiments against external baselines.

full rationale

The central CoATA pipeline (TEA + AIT + DPA) is an empirical method whose performance is evaluated on seven standard datasets against eleven external baselines using standard splits. No parameter is fitted to a quantity and then reported as a prediction of that same quantity; the augmented adjacency is constructed from features and topology, not from the test labels. The paper's self-citations (refs. [28,29]) are related-work references and are not used to justify any load-bearing premise. Theorem 4.1 is explicitly credited to the external GeoMix paper [60], and Theorem 4.2's proof is merely deferred, not circular. The only genuinely circular element is Theorem 4.3, which restates the contrastive loss definition as a 'guarantee.' This is a theoretical packaging issue rather than a circular derivation of the empirical claims. The skeptic's criticism that Algorithm 2 does not compute standard Personalized PageRank is a correctness/fidelity concern, not a circularity concern, and therefore does not raise the circularity score.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the assumption that PPR over an enriched node-attribute bipartite graph produces semantically useful edges, on the correctness of an unvalidated push-based PPR approximation, on an imported heterophily theorem, and on an unstated hyperparameter selection protocol. No new physical or ontological entities are introduced; all constructs are algorithmic constructions internal to the method.

free parameters (6)

alpha (PPR teleportation probability) = 0.2 (homophilic), 0.5 (heterophilic); selected per dataset from [0.1,0.9]
Controls restart probability in the bipartite PPR used to reconstruct edges; tuned per dataset in Section 5.5.
beta (residual mixing coefficient) = ~0.3 (homophilic), ~0.5 (heterophilic); from [0.1,0.9]
Controls how much of the original node attributes are retained during higher-order propagation in TEA (Eq. 4).
h (propagation depth) = 2 or 3 (from [1,4])
Number of propagation steps in Algorithm 1; affects how far attribute signals travel.
rmax (push threshold in Algorithm 2)
Threshold controlling PPR approximation precision versus runtime; value not reported in the main text.
K (KNN graph size) or k_add/k_del
Controls the number of edges added or removed in the co-augmented adjacency; chosen to keep edge count comparable to the original graph, but exact values are not reported.
lambda_1, lambda_2, lambda_3 (loss weights)
Weights for supervised, consistency, and prototype-alignment losses in Eq. (12); values are not reported.

assumptions (6)

domain assumption Heterophily alleviation via residual propagation (Theorem 4.1 from GeoMix [60]) holds for the evaluated graphs
This imported theorem is used to justify the TEA module (Eq. 4); the proof is omitted and assumed to carry over to the datasets used.
domain assumption Bipartite PPR over enriched features produces node similarities that reflect semantic relatedness
The AIT module (Section 4.2) assumes that high bipartite PPR scores between nodes correspond to useful, class-relevant edges; the paper provides no evidence beyond the final accuracy numbers.
domain assumption Algorithm 2's intra-set forward push correctly approximates the PPR distribution restricted to V
The residue flushing at attribute nodes is a nonstandard variant; its approximation error is not bounded in the main text and the proof is deferred to a supplementary not included in the arXiv version.
domain assumption Hyperparameters (alpha, beta, h, K, rmax, lambda) are selected on validation sets without test leakage
Section 5.1 reports search ranges but not a nested validation scheme; the reported test results implicitly assume no leakage during tuning.
standard math Standard public benchmark splits are used unmodified
Evaluation uses splits from [39], [41], and [35]; these are standard but not independently verified in this paper.
ad hoc to paper Minimizing the contrastive loss L_dpa yields cross-channel class-level alignment
Theorem 4.3 is a restatement of Eq. (11); its guarantee is by construction of the loss rather than an independent mathematical result.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoATA: Effective Co-Augmentation of Topology and Attribute for Graph Neural Networks." pith.science (2026). https://pith.science/paper/QZDPQJBE

@misc{pith2026250622299,
  author       = {Pith},
  title        = {Pith review of: CoATA: Effective Co-Augmentation of Topology and Attribute for Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QZDPQJBE}},
  note         = {Machine review of arXiv:2506.22299}
}

read the original abstract

Graph Neural Networks (GNNs) have garnered substantial attention due to their remarkable capability in learning graph representations. However, real-world graphs often exhibit substantial noise and incompleteness, which severely degrades the performance of GNNs. Existing methods typically address this issue through single-dimensional augmentation, focusing either on refining topology structures or perturbing node attributes, thereby overlooking the deeper interplays between the two. To bridge this gap, this paper presents CoATA, a dual-channel GNN framework specifically designed for the Co-Augmentation of Topology and Attribute. Specifically, CoATA first propagates structural signals to enrich and denoise node attributes. Then, it projects the enhanced attribute space into a node-attribute bipartite graph for further refinement or reconstruction of the underlying structure. Subsequently, CoATA introduces contrastive learning, leveraging prototype alignment and consistency constraints, to facilitate mutual corrections between the augmented and original graphs. Finally, extensive experiments on seven benchmark datasets demonstrate that the proposed CoATA outperforms eleven state-of-the-art baseline methods, showcasing its effectiveness in capturing the synergistic relationship between topology and attributes.

Figures

Figures reproduced from arXiv: 2506.22299 by the authors.

**Figure 2.** The framework of CoATA. (a) TEA improves node features via higher-order propagation; (b) AIT refines graph structure through a node-attribute bipartite graph. The model has two channels: (c) Original channel processes the raw graph; (d) Augmented channel leverages the co-augmented graph. Both share a GNN encoder, with DPA ensuring cross-view consistency. Supervised losses (𝐿𝑐𝑒 + 𝐿 ′ 𝑐𝑒 ) guide label predictions, and… view at source ↗

**Figure 3.** Nodes 𝑣1, 𝑣2, 𝑣3, 𝑣7 share the same class. AIT connects distant nodes (e.g., 𝑣7) via multi-hop feature walks. information integrated at each step. As indicated in Theorem 4.1, this residual mixing gradually strengthens the alignment with the correct class center, thereby mitigating feature contamination from dissimilar neighbors. This prevents oversmoothing, where nodes of different classes become too similar, but i… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Visualization of our model CoATA. Nodes are colored by their corresponding labels. can inadvertently reinforce mismatched neighbors. (2) CoATA outperforms all competitors by at least 1.31% and 1.40%, respectively, confirming that its dual-channel architecture effecti…

**Figure 5.** Figure 5: Hyperparameter Analysis. Left bar charts depict average accuracy and standard deviations for ℎ, highlighting best/worst results by green/red arrows. Right plots show how varying 𝛼 and 𝛽 affect performance under different ℎ. Teleportation Probability 𝛼. In our push-bas…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 56 canonical work pages

[1]

Reid Andersen, Fan R. K. Chung, and Kevin J. Lang. 2006. Local Graph Partitioning using PageRank Vectors. In FOCS

work page 2006
[2]

Mehdi Azabou, Venkataramana Ganesh, Shantanu Thakoor, Chi-Heng Lin, Lak- shmi Sathidevi, Ran Liu, Michal Valko, Petar Velickovic, and Eva L. Dyer. 2023. Half-Hop: A graph upsampling approach for slowing down message passing. In ICML

work page 2023
[3]

Aleksandar Bojchevski, Johannes Gasteiger, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek Rózemberczki, Michal Lukasik, and Stephan Günnemann. 2020. Scaling graph neural networks with approximate pagerank. In KDD

work page 2020
[4]

Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. 2020. Measuring and Relieving the Over-Smoothing Problem for Graph Neural Networks from the Topological View. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Sy...

work page 2020
[5]

Ming Chen, Zhewei Wei, Bolin Ding, Yaliang Li, Ye Yuan, Xiaoyong Du, and Ji-Rong Wen. 2020. Scalable graph neural networks via bidirectional propagation. Advances in neural information processing systems 33 (2020), 14556–14566

work page 2020
[6]

Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020. Simple and Deep Graph Convolutional Networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event

work page 2020
[7]

Yu Chen, Lingfei Wu, and Mohammed J. Zaki. 2020. Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

work page 2020
[8]

Kaize Ding, Yancheng Wang, Yingzhen Yang, and Huan Liu. 2023. Eliciting Structural and Semantic Global Knowledge in Unsupervised Graph Contrastive Learning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational A...

work page 2023

Show all 64 references

[9]

Kaize Ding, Zhe Xu, Hanghang Tong, and Huan Liu. 2022. Data Augmentation for Deep Graph Learning: A Survey. SIGKDD Explor. 24, 2 (2022)

2022
[10]

Chawla, and Huan Liu

Kaize Ding, Chuxu Zhang, Jie Tang, Nitesh V. Chawla, and Huan Liu. 2022. To- ward Graph Minimally-Supervised Learning. In KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, Au- gust 14 - 18, 2022

2022
[11]

Taoran Fang, Zhiqing Xiao, Chunping Wang, Jiarong Xu, Xuan Yang, and Yang Yang. 2023. DropMessage: Unifying Random Dropping for Graph Neural Net- works. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications ...

2023
[12]

Wenzheng Feng, Yuxiao Dong, Tinglin Huang, Ziqi Yin, Xu Cheng, Evgeny Khar- lamov, and Jie Tang. 2022. GRAND+: Scalable Graph Random Neural Networks. In WWW ’22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022

2022
[13]

Wenzheng Feng, Jie Zhang, Yuxiao Dong, Yu Han, Huanbo Luan, Qian Xu, Qiang Yang, Evgeny Kharlamov, and Jie Tang. 2020. Graph Random Neural Networks for Semi-Supervised Learning on Graphs. In NeurIPS

2020
[14]

Luca Franceschi, Mathias Niepert, Massimiliano Pontil, and Xiao He. 2019. Learn- ing Discrete Structures for Graph Neural Networks

2019
[15]

Hongyu Guo and Yongyi Mao. 2021. Intrusion-Free Graph Mixup. CoRR abs/2110.09344 (2021)

2021 arXiv
[16]

Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

2017
[17]

Xiaotian Han, Zhimeng Jiang, Ninghao Liu, and Xia Hu. 2022. G-Mixup: Graph Data Augmentation for Graph Classification

2022
[18]

Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In SIGIR. ACM, 639–648

2020
[19]

Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt. 1983. Sto- chastic blockmodels: First steps. Social networks 5, 2 (1983)

1983
[20]

Bo Jiang, Ziyan Zhang, Doudou Lin, Jin Tang, and Bin Luo. 2019. Semi-Supervised Learning With Graph Learning-Convolutional Networks. In 2019 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR)

2019
[21]

Wei Jin, Tyler Derr, Yiqi Wang, Yao Ma, Zitao Liu, and Jiliang Tang. 2021. Node Similarity Preserving Graph Convolutional Networks. In WSDM

2021
[22]

Nikhil Varma Keetha, Chen Wang, Yuheng Qiu, Kuan Xu, and Sebastian A. Scherer. 2022. AirObject: A Temporally Evolving Graph Embedding for Object Identification. In CVPR

2022
[23]

Kipf and Max Welling

Thomas N. Kipf and Max Welling. 2016. Variational Graph Auto-Encoders. CoRR abs/1611.07308 (2016)

2016 arXiv
[24]

Kipf and Max Welling

Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In ICLR

2017
[25]

Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. 2019. Pre- dict then Propagate: Graph Neural Networks meet Personalized PageRank. In ICLR

2019
[26]

Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. 2019. Pre- dict then Propagate: Graph Neural Networks meet Personalized PageRank. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

2019
[27]

Xiang Li, Renyu Zhu, Yao Cheng, Caihua Shan, Siqiang Luo, Dongsheng Li, and Weining Qian. 2022. Finding Global Homophily in Graph Neural Networks When Meeting Heterophily. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA

2022
[28]

Ziyu Liao, Tao Liu, Yue He, and Longlong Lin. 2024. Effective Temporal Graph Learning via Personalized PageRank. Entropy 26, 7 (2024), 588

2024
[29]

Longlong Lin, Yunfeng Yu, Zihao Wang, Zeli Wang, Yuying Zhao, Jin Zhao, and Tao Jia. 2024. PSNE: Efficient Spectral Sparsification Algorithms for Scaling Network Embedding. In CIKM. 1420–1429

2024
[30]

Yixin Liu, Kaize Ding, Jianling Wang, Vincent C. S. Lee, Huan Liu, and Shirui Pan. 2023. Learning Strong Graph Neural Networks with Weak Information. In KDD 2023, Long Beach, CA, USA, August 6-10, 2023

2023
[31]

Sitao Luan, Chenqing Hua, Qincheng Lu, Jiaqi Zhu, Mingde Zhao, Shuyuan Zhang, Xiao-Wen Chang, and Doina Precup. 2022. Revisiting Heterophily For Graph Neural Networks. In Advances in Neural Information Processing Systems 35:Annual Conference on Neural Information Processing Sy...

2022
[32]

Dongsheng Luo, Wei Cheng, Wenchao Yu, Bo Zong, Jingchao Ni, Haifeng Chen, and Xiang Zhang. 2021. Learning to Drop: Robust Graph Neural Network via Topological Denoising. In WSDM

2021
[33]

Yuchen Meng, Ronghua Li, Longlong Lin, Xunkai Li, and Guoren Wang. 2024. Topology-preserving Graph Coarsening: An Elementary Collapse-based Ap- proach. Proc. VLDB Endow. 17, 13 (2024), 4760–4772

2024
[34]

Hyeon-Jin Park, Seunghun Lee, Sihyeon Kim, Jinyoung Park, Jisu Jeong, Kyung- Min Kim, Jung-Woo Ha, and Hyunwoo J. Kim. 2021. Metropolis-Hastings Data Augmentation for Graph Neural Networks. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural In...

2021
[35]

Oleg Platonov, Denis Kuznedelev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. 2023. A critical look at the evaluation of GNNs under heterophily: Are we really making progress? (2023)

2023
[36]

Jiezhong Qiu, Jian Tang, Hao Ma, Yuxiao Dong, Kuansan Wang, and Jie Tang
[37]

Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. 2020. DropE- dge: Towards Deep Graph Convolutional Networks on Node Classification. In International Conference on Learning Representations

2020
[38]

Constantin Marc Seibold, Simon Reiß, Jens Kleesiek, and Rainer Stiefelhagen. 2022. Reference-Guided Pseudo-Label Generation for Medical Semantic Segmentation. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty- Fourth Conference on Innovative Applica...

2022
[39]

Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi-Rad. 2008. Collective Classification in Network Data. AI Mag. 29, 3 (2008)

2008
[40]

Chao Shang, Jie Chen, and Jinbo Bi. 2021. Discrete Graph Structure Learning for Forecasting Multiple Time Series. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021

2021
[41]

Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of Graph Neural Network Evaluation. CoRR (2018)

2018
[42]

Dyer, Rémi Munos, Petar Velickovic, and Michal Valko

Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Mehdi Azabou, Eva L. Dyer, Rémi Munos, Petar Velickovic, and Michal Valko. 2022. Large-Scale Representation Learning on Graphs via Bootstrapping. In The Tenth International Conference on Learning Representations, ICL...

2022
[43]

Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In International Con- ference on Learning Representations

2018
[44]

Vikas Verma, Meng Qu, Kenji Kawaguchi, Alex Lamb, Yoshua Bengio, Juho Kannala, and Jian Tang. 2021. GraphMix: Improved Training of GNNs for Semi- Supervised Learning. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative A...

2021
[45]

Ruijia Wang, Shuai Mou, Xiao Wang, Wanpeng Xiao, Qi Ju, Chuan Shi, and Xing Xie. 2021. Graph Structure Estimation Neural Networks. In WWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021

2021
[46]

Aggarwal, Yi Chang, and Huan Liu

Suhang Wang, Jiliang Tang, Charu C. Aggarwal, Yi Chang, and Huan Liu. 2017. Signed Network Embedding in Social Media. In SIAM. CoATA: Effective Co-Augmentation of Topology and Attribute for Graph Neural Networks ICMR ’25, June 30-July 3, 2025, Chicago, IL, USA

2017
[47]

Tongzhou Wang and Phillip Isola. 2020. Understanding Contrastive Represen- tation Learning through Alignment and Uniformity on the Hypersphere. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event

2020
[48]

Yiwei Wang, Wei Wang, Yuxuan Liang, Yujun Cai, and Bryan Hooi. 2021. Mixup for Node and Graph Classification. InWWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021

2021
[49]

Zeli Wang, Jian Li, Shuyin Xia, Longlong Lin, and Guoyin Wang. 2024. Text Adversarial Defense via Granular-Ball Sample Enhancement. In ICMR. 348–356

2024
[50]

Zeli Wang, Tuo Zhang, Shuyin Xia, Longlong Lin, and Guoyin Wang. 2024. GBRAIN: Combating Textual Label Noise by Granular-ball based Robust Training. In ICMR. 357–365

2024
[51]

Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying Graph Convolutional Networks. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA

2019
[52]

Jun Xia, Lirong Wu, Ge Wang, Jintao Chen, and Stan Z. Li. 2022. ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning. InInternational Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA

2022
[53]

Guotong Xue, Ming Zhong, Tieyun Qian, and Jianxin Li. 2024. PSA-GNN: An augmented GNN framework with priori subgraph knowledge. Neural Networks 173 (2024)

2024
[54]

Liang Yang, Zesheng Kang, Xiaochun Cao, Di Jin, Bo Yang, and Yuanfang Guo
[55]

Hamilton, and Jure Leskovec

Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2018. Graph Convolutional Neural Networks for Web-Scale Recommender Systems. In SIGKDD

2018
[56]

Yunfeng Yu, Longlong Lin, Qiyu Liu, Zeli Wang, Xi Ou, and Tao Jia. 2024. GSD- GNN: Generalizable and Scalable Algorithms for Decoupled Graph Neural Net- works. In ICMR. 64–72

2024
[57]

Han Zhao, Xu Yang, Zhenru Wang, Erkun Yang, and Cheng Deng. 2021. Graph Debiased Contrastive Learning with Joint Representation Clustering. In Proceed- ings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada...

2021
[58]

Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N. Metaxas. 2019. Semantic Graph Convolutional Networks for 3D Human Pose Regression. In CVPR

2019
[59]

Tong Zhao, Yozen Liu, Leonardo Neves, Oliver Woodford, Meng Jiang, and Neil Shah. 2021. Data Augmentation for Graph Neural Networks

2021
[60]

Wentao Zhao, Qitian Wu, Chenxiao Yang, and Junchi Yan. 2024. GeoMix: Towards Geometry-Aware Data Augmentation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024

2024
[61]

Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. Deep Graph Contrastive Representation Learning. CoRR abs/2006.04131 (2020)

2020 arXiv
[62]

Chenyi Zhuang and Qiang Ma. 2018. Dual Graph Convolutional Networks for Graph-Based Semi-Supervised Classification. In Proceedings of the 2018 World Wide Web Conference on World Wide Web, WWW 2018, Lyon, France, April 23-27, 2018

2018
[2018]

In SIGKDD

DeepInf: Social Influence Prediction with Deep Learning. In SIGKDD
[2019]

InProceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019

Topology Optimization based Graph Convolutional Network. InProceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019

2019

Pith tools

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

[1] [1]

Reid Andersen, Fan R. K. Chung, and Kevin J. Lang. 2006. Local Graph Partitioning using PageRank Vectors. In FOCS

work page 2006

[2] [2]

Mehdi Azabou, Venkataramana Ganesh, Shantanu Thakoor, Chi-Heng Lin, Lak- shmi Sathidevi, Ran Liu, Michal Valko, Petar Velickovic, and Eva L. Dyer. 2023. Half-Hop: A graph upsampling approach for slowing down message passing. In ICML

work page 2023

[3] [3]

Aleksandar Bojchevski, Johannes Gasteiger, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek Rózemberczki, Michal Lukasik, and Stephan Günnemann. 2020. Scaling graph neural networks with approximate pagerank. In KDD

work page 2020

[4] [4]

Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. 2020. Measuring and Relieving the Over-Smoothing Problem for Graph Neural Networks from the Topological View. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Sy...

work page 2020

[5] [5]

Ming Chen, Zhewei Wei, Bolin Ding, Yaliang Li, Ye Yuan, Xiaoyong Du, and Ji-Rong Wen. 2020. Scalable graph neural networks via bidirectional propagation. Advances in neural information processing systems 33 (2020), 14556–14566

work page 2020

[6] [6]

Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020. Simple and Deep Graph Convolutional Networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event

work page 2020

[7] [7]

Yu Chen, Lingfei Wu, and Mohammed J. Zaki. 2020. Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

work page 2020

[8] [8]

Kaize Ding, Yancheng Wang, Yingzhen Yang, and Huan Liu. 2023. Eliciting Structural and Semantic Global Knowledge in Unsupervised Graph Contrastive Learning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational A...

work page 2023