Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Catch Causal Signals from Edges for Label Imbalance in Graph Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Edge features routed through causal attention improve graph classification under label imbalance.

desk verdict Useful edge-feature extension of CAL, but the label-imbalance protocol is so under-specified that the exact numbers are not yet reproducible. read the letter →

arxiv 2501.01707 v2 pith:R7ZBPSFT submitted 2025-01-03 cs.LG

Fengrui Zhang , Yujia Yin , Hongzong Li , Yifan Chen , Tianyi Qu This is my paper · ORCID
classification cs.LG
keywords graphclassificationlabelimbalancecausalattentionedgefeaturesout-of-distributiongeneralizationneuralnetworkssubgraphbackdooradjustment
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 sets out to show that graph edge features carry causal information that node-only attention misses, and that wiring edge features into the causal attention mechanism improves graph classification when training and test label distributions differ. It introduces ECAL, which replaces the node-only attention of the CAL framework with edge-featured EGATv1/EGATv2 encoders, splits each graph into causal and trivial subgraphs using node and edge attention masks, and trains with a three-part loss. Across PTC, Tox21, and ogbg-molhiv, ECAL+EGATv2 reports the best accuracy on most tasks, including 0.671 versus 0.429 on PTC-FM and 0.997 on Tox21-PPAR-gamma, with p-values below 0.05 in all comparisons. A sympathetic reader would care because label imbalance is common in molecular and toxicity prediction, and this is a direct upgrade to an existing out-of-distribution method rather than a new architecture.

What carries the argument

The load-bearing mechanism is EGATv2 embedded in the causal attention pipeline. EGATv2 first computes an edge embedding $e^{\text{out}}_{ij} = \mathrm{LeakyReLU}(W^{\mathrm{out}}_e [X'_i \| e'_{ij} \| X'_j])$ from the projected node and edge features, then uses that edge embedding both in the attention weights $a_{ij}$ for neighbor aggregation and as input to MLPs that output node and edge causality scores. Those scores become diagonal masks that partition the original graph into a causal subgraph and a trivial subgraph, and the three-part loss (cross-entropy on causal, KL-to-uniform on trivial, cross-entropy on randomly paired causal and trivial representations) then shapes the representation. This design makes edge information do work at every stage: detecting causality, splitting the graph, and constructing representations.

What would settle it

Run ECAL and CAL on the same benchmarks with the label-imbalance protocol fully specified (imbalance ratio, sampling method, seeds) and with datasets that have natural, documented label imbalance; if the edge-enhanced model's accuracy advantage disappears or reverses when the shift is controlled, the central claim would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that edge features are a primary source of causal signal in graphs, and that a causal attention mechanism which estimates both node-level and edge-level causality scores can separate the causal subgraph from the trivial subgraph better than node-only attention, thereby improving out-of-distribution accuracy under label imbalance. Concretely, the paper modifies the CAL framework so that the attention estimation step uses EGATv2's edge feature evolution: edge embeddings are formed from concatenated node and edge projections, then used both to aggregate neighbor features and to score node and edge causality $\alpha_i$ and $\alpha_{ij}$. These scores act as masks $M_x$, $M_e$ that split the graph into $G_c$ and $G_t$, and both subgraphs are encoded with edge-featured EGATs before a readout and classifier. The paper's reported finding is that ECAL+EGATv2 outperforms CAL+GCN, CAL+GAT, and plain GCN/GAT/EGAT baselines on most tasks, and that ablating the KL-uniformity or backdoor-adjustment loss terms hurts ECAL while barely affecting CAL, supporting the claim that the edge-enhanced causal detection terms carry the improvement.

Load-bearing premise

The load-bearing premise is that the paper's undocumented manipulation of the datasets actually creates a train/test distribution shift that reflects label imbalance in the real world; if that manipulation is arbitrary, the reported edge-feature gains may be an artifact of the specific split.

Editorial extensions

If this is right

  • On PTC-FM, ECAL+EGATv2 reaches 0.671 test accuracy versus 0.429 for CAL+GCN and CAL+GAT, a jump of more than 24 percentage points on that task.
  • On ogbg-molhiv, ECAL+EGATv2 reaches 0.81, the best result in the table, and ten-trial t-tests report p-values as low as 7.3e-14 against CAL.
  • Randomly permuting an increasing fraction of edge feature vectors decreases ECAL's out-of-distribution accuracy on ogbg-molhiv, which the paper reads as evidence that edge features are causally relevant rather than peripheral.
  • Ablating the KL divergence or backdoor-adjustment terms slightly changes CAL's accuracy but clearly lowers ECAL's, indicating the causality detection loss becomes effective once edge features are included.

Reading between the lines

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

  • A testable extension is to run ECAL on graphs whose edges are merely topological, without bond or interaction attributes; the paper's logic predicts smaller gains there because the edge features carry less task-relevant signal.
  • Since the paper does not specify imbalance ratios or the sampling protocol, comparing ECAL and CAL under a fully specified sweep of imbalance levels would show whether the reported margin is stable or concentrated at certain shift strengths.
  • The noise-permutation ablation could be paired with an equivalent node-feature permutation ablation to test whether edge features are causal in a stronger sense than node features.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes ECAL, an edge-enhanced causal attention learning framework for graph classification under label imbalance. It extends the CAL causal-attention framework by incorporating edge-featured graph attention (EGATv1/v2) into both attention estimation and graph representation steps, splitting each graph into a causal and a trivial subgraph via learned node/edge attention scores, and training with a combination of cross-entropy, KL-divergence-to-uniform, and backdoor-adjustment losses. The authors report test accuracies on PTC-FM, PTC-MM, several Tox21 endpoints, and ogbg-molhiv, claiming consistent improvements over GCN, GAT, EGAT, and CAL baselines under a label-imbalance manipulation described only as making training and testing distributions different.

Significance. If the reported results hold, the empirical contribution is moderately valuable: adding edge-feature awareness to causal-attention graph classifiers appears to improve predictive accuracy under label-imbalanced splits, and the effect sizes on PTC-FM and PTC-MM are substantial. Strengths of the paper include the public code repository, the reporting of p-values from 10 independent trials for ECAL-versus-CAL comparisons, and the inclusion of two ablation studies (removing causality-detection losses and permuting edge features). However, the significance is currently limited by the absence of a reproducible description of the label-imbalance protocol, which is the very condition under which all comparisons are made. The causal interpretation of the learned attention masks is also not independently validated, so the paper is best viewed as an empirical method comparison until the protocol and validation details are supplied.

major comments (4)
  1. [Section IV, first paragraph] The label-imbalance manipulation is never specified. The text states only that "We manipulate the dataset to introduce label imbalance so that the distribution of training datasets and that of testing datasets are different," with no imbalance ratios, sampling mechanism, or splitting rule for any of the seven reported tasks. Because every number in Table I is defined relative to this undisclosed manipulation, the central claim that ECAL improves OOD accuracy under label imbalance is not reproducible, and the reported margins (e.g., 0.671 vs. 0.429 on PTC-FM) could be artifacts of an arbitrary or seed-favorable split.
  2. [Table I and Table II] Table I reports only point estimates of test accuracy, and Table II gives p-values for ECAL versus CAL but no standard deviations, confidence intervals, or per-trial results. The p-values indicate that the ECAL-vs-CAL difference is unlikely to be zero, but without the variance and effect size across the 10 trials, the reader cannot assess the magnitude or stability of the improvement, and no significance test is provided against the non-CAL baselines that ECAL is also claimed to beat.
  3. [Section IV-B and Figure 3] The ablation that permutes edge features is under-specified: the three "High/Middle/Low Imbalance" levels are not defined, the proportion of permuted edge feature vectors is the only varying quantity, and the permutation procedure (within a graph, across graphs, or over the whole dataset) is not described. Without these details, the conclusion that OOD accuracy decreases because edge features are causal is not established; the observed degradation could depend on the specific noise injection scheme.
  4. [Section III-B and III-B.3] The "causal" label for the subgraph is self-referential in a way that weakens the causal claim. The attention scores α_i and α_ij that define the causal subgraph are learned using the training labels through the cross-entropy loss on the causal subgraph (Eq. 8), and then the same labels are used to validate that this subgraph captures causality. The paper should either temper the causal interpretation to "label-relevant attention" or provide an independent validation, such as an intervention-style test or an out-of-sample criterion that does not use the training labels to select the causal subgraph.
minor comments (5)
  1. [Figure 2] The x-axis of Figure 2 includes Tox21-AhR, but this endpoint does not appear in Table I; please clarify whether the results come from a different experimental setting or whether the figure should be restricted to the reported endpoints.
  2. [Section III-A, Eq. (1)-(4)] The symbol a is used both for the learnable attention vector in the softmax computations and for the scalar attention score a_ij; please use distinct notation to avoid confusion.
  3. [Equation (11)] The hyperparameters λ1 and λ2 are never specified in the experiments; please report their values and, if possible, a sensitivity analysis, since the loss balance directly affects the trade-off between the causal and trivial subgraph losses.
  4. [Reference [30]] Reference [30] is cited for the EGATv2 design, but the cited work appears to be the Rossmann-toolbox bioinformatics protocol rather than a graph attention network; please verify and correct the citation.
  5. [Abstract and Figure 1] The abstract contains the typo "real-word datasets" (should be "real-world"), and Figure 1 labels the graph as "Trival Attended graph" (should be "Trivial Attended graph"); these can be corrected in a final polish.

Circularity Check

1 steps flagged · score 3.0 of 10

No load-bearing circularity in the central comparison; one self-definitional noise test in the edge-causality ablation and an under-specified label-imbalance protocol.

  1. self definitional [Section IV-B.2 (Ablation: adding noise to ablate edge features) and Figure 3]
    "In causal learning, features are considered non-causal if adding noise does not reduce OOD accuracy. To emphasize the causality of edge features, we randomly permute a proportion of edge feature vectors. As shown in Figure 3, adding noise consistently reduces OOD accuracy on ogbg-molhiv, confirming the causal significance of edge features."

    The paper's criterion for causal significance is the same noise-permutation test used as evidence: a feature is labeled non-causal if adding noise does not reduce accuracy, so observing that noise reduces accuracy is a restatement of the operational definition rather than an independent confirmation. The claim that edge features are causal follows from the test by construction. This circularity is secondary: the main benchmark comparison (ECAL vs CAL/GCN/GAT in Table I) is an external accuracy comparison and does not reduce to this definitional loop.

full rationale

The central claim is an empirical comparison, not a derived prediction: ECAL+EGATv2 is evaluated against GCN, GAT, EGATv1, EGATv2, CAL+GCN, and CAL+GAT on standard benchmarks, and the reported margins in Table I are external test-accuracy facts rather than quantities implied by the model's own equations. The causal-subgraph construction (Eqs. 5-6) uses attention masks learned through the classification loss, which is a standard supervised masking mechanism, not a self-referential prediction of the same quantity used to fit it. The one self-citation, [21] by co-author Yifan Chen, is cited only as an example of mutual-information causal-component identification in the related-work survey and is not load-bearing for the proposed method or results. The main reproducibility limitation is the one-sentence label-imbalance manipulation in Sec. IV ('We manipulate the dataset to introduce label imbalance so that the distribution of training datasets and that of testing datasets are different'), which omits imbalance ratios, sampling mechanisms, and split rules; this is a missing experimental specification, not circularity. Overall, the derivation chain is self-contained: the EGAT modules are explicitly attributed to prior work, the causal-attention framework is inherited from CAL, and the paper's contribution is a benchmarked integration of edge features, except for the secondary noise-test interpretative step noted above.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical or formal entities; causal and trivial subgraphs are algorithmic constructs derived from learned masks. The core assumptions are the causal validity of attention masks, the backdoor-adjustment interpretation of random pairing, and the representativeness of the undisclosed imbalance protocol.

free parameters (2)
  • Label-imbalance ratios and train/test shift protocol = not reported
    The central comparison depends on how each dataset was imbalanced; the paper does not specify the ratios or the train/test shift mechanism.
  • Loss weights lambda_1 and lambda_2 = not reported
    Equation (11) introduces two scalar weights that control the causality-detection losses; no values or tuning procedure are given, and ablations show these terms matter.
assumptions (3)
  • domain assumption Learned node and edge attention scores alpha_i and alpha_ij identify true causal structure
    No causal identifiability conditions or ground-truth causal subgraphs are provided; the causal interpretation is imposed in Sections III-A and III-B.
  • domain assumption Random pairing of causal and trivial representations implements backdoor adjustment
    Equation (10) assumes h_c and h_t are additively separable components such that pairing approximates intervention; this is taken from cited prior work without justification for graph representations.
  • domain assumption The induced label-imbalance shift is a representative OOD evaluation
    Section IV does not describe the manipulation; all conclusions about OOD generalization rest on this unstated protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Catch Causal Signals from Edges for Label Imbalance in Graph Classification." pith.science (2026). https://pith.science/paper/R7ZBPSFT

@misc{pith2026250101707,
  author       = {Pith},
  title        = {Pith review of: Catch Causal Signals from Edges for Label Imbalance in Graph Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7ZBPSFT}},
  note         = {Machine review of arXiv:2501.01707}
}
read the original abstract

Despite significant advancements in causal research on graphs and its application to cracking label imbalance, the role of edge features in detecting the causal effects within graphs has been largely overlooked, leaving existing methods with untapped potential for further performance gains. In this paper, we enhance the causal attention mechanism through effectively leveraging edge information to disentangle the causal subgraph from the original graph, as well as further utilizing edge features to reshape graph representations. Capturing more comprehensive causal signals, our design leads to improved performance on graph classification tasks with label imbalance issues. We evaluate our approach on real-word datasets PTC, Tox21, and ogbg-molhiv, observing improvements over baselines. Overall, we highlight the importance of edge features in graph causal detection and provide a promising direction for addressing label imbalance challenges in graph-level tasks. The model implementation details and the codes are available on https://github.com/fengrui-z/ECAL

Figures

Figures reproduced from arXiv: 2501.01707 by the authors.

Figure 1
Figure 1. Framework overview. The proposed model integrates causal attention [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The effect on OOD accuracy of removing causality detection loss [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. OOD accuracy under different noise levels (proportion of permutation) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 26 canonical work pages

  1. [1]

    The graph neural network model,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfar- dini, “The graph neural network model,” IEEE transactions on neural networks, vol. 20, no. 1, pp. 61–80, 2008

    2008

  2. [2]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and S. Y . Philip, “A comprehensive survey on graph neural networks,” IEEE transactions on neural networks and learning systems , vol. 32, no. 1, pp. 4–24, 2020

    2020

  3. [3]

    Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking,

    Z. Wu, J. Wang, H. Du, J. dejun, Y . Kang, D. Li, P. Pan, Y . Deng, D.-S. Cao, K. Hsieh, and T. Hou, “Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking,” Nature Communications, vol. 14, 05 2023

    work page 2023

  4. [4]

    Analyzing learned molecular representations for property prediction,

    K. Yang, K. Swanson, W. Jin, C. Coley, P. Eiden, H. Gao, A. Guzman- Perez, T. Hopper, B. Kelley, M. Mathea et al. , “Analyzing learned molecular representations for property prediction,” Journal of chemical information and modeling , vol. 59, no. 8, pp. 3370–3388, 2019

    work page 2019

  5. [5]

    Convolutional networks on graphs for learning molecular fingerprints,

    D. K. Duvenaud, D. Maclaurin, J. Iparraguirre, R. Bombarell, T. Hirzel, A. Aspuru-Guzik, and R. P. Adams, “Convolutional networks on graphs for learning molecular fingerprints,” Advances in neural information processing systems, vol. 28, 2015

    2015

  6. [6]

    Deeprank-gnn: a graph neural network framework to learn patterns in protein–protein interfaces,

    M. R ´eau, N. Renaud, L. C. Xue, and A. M. Bonvin, “Deeprank-gnn: a graph neural network framework to learn patterns in protein–protein interfaces,” Bioinformatics, vol. 39, no. 1, p. btac759, 2023

    work page 2023

  7. [7]

    Prediction of protein–protein interaction using graph neural networks,

    K. Jha, S. Saha, and H. Singh, “Prediction of protein–protein interaction using graph neural networks,” Scientific Reports, vol. 12, no. 1, p. 8360, 2022

    work page 2022

  8. [8]

    Imbalanced graph classification via graph-of-graph neural networks,

    Y . Wang, Y . Zhao, N. Shah, and T. Derr, “Imbalanced graph classification via graph-of-graph neural networks,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management , 2022, pp. 2067–2076

    work page 2022

Show all 39 references
  1. [9]

    Graphsmote: Imbalanced node classification on graphs with graph neural networks,

    T. Zhao, X. Zhang, and S. Wang, “Graphsmote: Imbalanced node classification on graphs with graph neural networks,” in Proceedings of the 14th ACM international conference on web search and data mining , 2021, pp. 833–841

    2021

  2. [10]

    Multi-class imbalanced graph convolutional network learning,

    M. Shi, Y . Tang, X. Zhu, D. Wilson, and J. Liu, “Multi-class imbalanced graph convolutional network learning,” in Proceedings of the Twenty- Ninth International Joint Conference on Artificial Intelligence (IJCAI- 20), 2020

    2020

  3. [11]

    Ins-gnn: Improving graph imbalance learning with self-supervision,

    X. Juan, F. Zhou, W. Wang, W. Jin, J. Tang, and X. Wang, “Ins-gnn: Improving graph imbalance learning with self-supervision,” Information Sciences, vol. 637, p. 118935, 2023

    2023

  4. [12]

    Good: A graph out-of-distribution benchmark,

    S. Gui, X. Li, L. Wang, and S. Ji, “Good: A graph out-of-distribution benchmark,” Advances in Neural Information Processing Systems , vol. 35, pp. 2059–2073, 2022

    2022

  5. [13]

    A survey of graph neural networks in real world: Imbalance, noise, privacy and ood challenges,

    W. Ju, S. Yi, Y . Wang, Z. Xiao, Z. Mao, H. Li, Y . Gu, Y . Qin, N. Yin, S. Wang et al. , “A survey of graph neural networks in real world: Imbalance, noise, privacy and ood challenges,” arXiv preprint arXiv:2403.04468, 2024

    2024

  6. [14]

    Recent advances in reliable deep graph learning: Inherent noise, distri- bution shift, and adversarial attack,

    J. Li, B. Wu, C. Hou, G. Fu, Y . Bian, L. Chen, J. Huang, and Z. Zheng, “Recent advances in reliable deep graph learning: Inherent noise, distri- bution shift, and adversarial attack,” arXiv preprint arXiv:2202.07114 , 2022

    2022 arXiv

  7. [15]

    Trustworthy graph neural networks: aspects, methods, and trends,

    H. Zhang, B. Wu, X. Yuan, S. Pan, H. Tong, and J. Pei, “Trustworthy graph neural networks: aspects, methods, and trends,” Proceedings of the IEEE, 2024

    2024

  8. [16]

    Learning substructure invariance for out-of-distribution molecular representations,

    N. Yang, K. Zeng, Q. Wu, X. Jia, and J. Yan, “Learning substructure invariance for out-of-distribution molecular representations,” Advances in Neural Information Processing Systems , vol. 35, pp. 12 964–12 978, 2022

    2022

  9. [17]

    Environment-aware dynamic graph learning for out-of-distribution generalization,

    H. Yuan, Q. Sun, X. Fu, Z. Zhang, C. Ji, H. Peng, and J. Li, “Environment-aware dynamic graph learning for out-of-distribution generalization,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  10. [18]

    Joint learning of label and environment causal independence for graph out-of-distribution general- ization,

    S. Gui, M. Liu, X. Li, Y . Luo, and S. Ji, “Joint learning of label and environment causal independence for graph out-of-distribution general- ization,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  11. [19]

    Out-of-distribution generalized dynamic graph neural network with disentangled interven- tion and invariance promotion,

    Z. Zhang, X. Wang, Z. Zhang, H. Li, and W. Zhu, “Out-of-distribution generalized dynamic graph neural network with disentangled interven- tion and invariance promotion,” arXiv preprint arXiv:2311.14255, 2023

    2023 arXiv

  12. [20]

    Towards human-like perception: Learning structural causal model in heterogeneous graph,

    T. Lin, K. Song, Z. Jiang, Y . Kang, W. Yuan, X. Li, C. Sun, C. Huang, and X. Liu, “Towards human-like perception: Learning structural causal model in heterogeneous graph,” Information Processing & Management, vol. 61, no. 2, p. 103600, 2024

    2024

  13. [21]

    Learning causally invariant representations for out-of- distribution generalization on graphs,

    Y . Chen, Y . Zhang, Y . Bian, H. Yang, M. Kaili, B. Xie, T. Liu, B. Han, and J. Cheng, “Learning causally invariant representations for out-of- distribution generalization on graphs,” Advances in Neural Information Processing Systems, vol. 35, pp. 22 131–22 148, 2022

    2022

  14. [22]

    Generalizing graph neural networks on out-of-distribution graphs,

    S. Fan, X. Wang, C. Shi, P. Cui, and B. Wang, “Generalizing graph neural networks on out-of-distribution graphs,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023

    2023

  15. [23]

    Causal discovery with attention- based convolutional neural networks,

    M. Nauta, D. Bucur, and C. Seifert, “Causal discovery with attention- based convolutional neural networks,” Machine Learning and Knowl- edge Extraction, vol. 1, no. 1, p. 19, 2019

    2019

  16. [24]

    Causal attention for interpretable and generalizable graph classification,

    Y . Sui, X. Wang, J. Wu, M. Lin, X. He, and T.-S. Chua, “Causal attention for interpretable and generalizable graph classification,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 1696–1705

    2022

  17. [25]

    Exploiting edge features for graph neural networks,

    L. Gong and Q. Cheng, “Exploiting edge features for graph neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 9211–9219

    2019

  18. [26]

    Strategies for pre-training graph neural networks,

    W. Hu, B. Liu, J. Gomes, M. Zitnik, P. Liang, V . Pande, and J. Leskovec, “Strategies for pre-training graph neural networks,” arXiv preprint arXiv:1905.12265, 2019

    1905 arXiv

  19. [27]

    Egat: Edge-featured graph attention network,

    Z. Wang, J. Chen, and H. Chen, “Egat: Edge-featured graph attention network,” in Artificial Neural Networks and Machine Learning–ICANN 2021: 30th International Conference on Artificial Neural Networks, Bratislava, Slovakia, September 14–17, 2021, Proceedings, Part I 30 . Sprin...

    2021

  20. [28]

    Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network,

    X. Mo, Z. Huang, Y . Xing, and C. Lv, “Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 7, pp. 9554–9567, 2022

    2022

  21. [29]

    Do transformers really perform badly for graph representation?

    C. Ying, T. Cai, S. Luo, S. Zheng, G. Ke, D. He, Y . Shen, and T.-Y . Liu, “Do transformers really perform badly for graph representation?” Advances in neural information processing systems, vol. 34, pp. 28 877– 28 888, 2021

    2021

  22. [30]

    Rossmann-toolbox: a deep learning-based protocol for the prediction and design of cofactor speci- ficity in rossmann fold proteins,

    K. Kami ´nski, J. Ludwiczak, M. Jasi ´nski, A. Bukala, R. Madaj, K. Szczepaniak, and S. Dunin-Horkawicz, “Rossmann-toolbox: a deep learning-based protocol for the prediction and design of cofactor speci- ficity in rossmann fold proteins,” Briefings in bioinformatics , vol. 23,...

    2022

  23. [31]

    Pearl, Causality

    J. Pearl, Causality. Cambridge university press, 2009

    2009

  24. [32]

    Causal effect identification by adjustment under confounding and selection biases,

    J. Correa and E. Bareinboim, “Causal effect identification by adjustment under confounding and selection biases,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 31, no. 1, 2017

    2017

  25. [33]

    Generalized adjustment under confounding and selection biases,

    J. Correa, J. Tian, and E. Bareinboim, “Generalized adjustment under confounding and selection biases,” in Proceedings of the AAAI Confer- ence on Artificial Intelligence , vol. 32, no. 1, 2018

    2018

  26. [34]

    Deeptox: toxicity prediction using deep learning,

    A. Mayr, G. Klambauer, T. Unterthiner, and S. Hochreiter, “Deeptox: toxicity prediction using deep learning,” Frontiers in Environmental Science, vol. 3, p. 80, 2016

    2016

  27. [35]

    Tox21challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and drugs,

    R. Huang, M. Xia, D.-T. Nguyen, T. Zhao, S. Sakamuru, J. Zhao, S. A. Shahane, A. Rossoshek, and A. Simeonov, “Tox21challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and drugs,” Frontiers in ...

    2016

  28. [36]

    Open graph benchmark: Datasets for machine learning on graphs,

    W. Hu, M. Fey, M. Zitnik, Y . Dong, H. Ren, B. Liu, M. Catasta, and J. Leskovec, “Open graph benchmark: Datasets for machine learning on graphs,” Advances in neural information processing systems , vol. 33, pp. 22 118–22 133, 2020

    2020

  29. [37]

    Unsupervised inductive graph-level representation learning via graph-graph proximity,

    Y . Bai, H. Ding, Y . Qiao, A. Marinovic, K. Gu, T. Chen, Y . Sun, and W. Wang, “Unsupervised inductive graph-level representation learning via graph-graph proximity,” arXiv preprint arXiv:1904.01098 , 2019

    1904 arXiv

  30. [38]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  31. [39]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.