Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

A directed-graph contrastive learning framework that jointly perturbs the magnetic Laplacian in the complex domain and samples local paths in the real domain claims state-of-the-art accuracy on seven digraph benchmarks without label supervi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 09:14 UTC pith:HC333KBB

load-bearing objection A solid digraph GCL paper with real gains, but the headline numbers don't match its own tables and Theorem 2 needs a corrected proof. the 3 major comments →

arxiv 2510.16311 v3 pith:HC333KBB submitted 2025-10-18 cs.LG

Toward General Digraph Contrastive Learning: A Dual Spatial Perspective

Zhengyu Wu , Daohan Su , Yang Zhang , Xunkai Li , Rong-Hua Li , Guoren Wang This is my paper
classification cs.LG
keywords directed graph contrastive learningmagnetic Laplacianpersonalized perturbationpath-based augmentationself-supervised graph learningdigraph representation learningVon Neumann entropynode classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that contrastive learning on directed graphs fails when augmentation ignores directionality, and it proposes a dual-spatial strategy to fix this: a complex-domain branch that perturbs the magnetic Laplacian with node-specific charge values, plus a real-domain branch that samples BFS/DFS paths and encodes them with a GRU. The two views are fused and trained with an InfoNCE objective. The authors report that this method, S2-DiGCL, consistently achieves the highest node classification accuracy across all seven datasets and the best link prediction results on all four datasets tested, improving over strong baselines by 4.41% and 4.34%, respectively, while matching or exceeding several supervised GNNs without using labels. The paper also offers a theoretical account via magnetic Von Neumann entropy, claiming that the perturbation produces monotone and bounded spectral change.

Core claim

S2-DiGCL claims that effective contrastive learning on directed graphs requires two complementary spatial perspectives: a complex-domain view where the magnetic Laplacian's edge phases are adaptively modulated by a topology-aware charge, and a real-domain view where direction-aware BFS/DFS paths around each node are sequentially encoded. Because the complex view captures global directional asymmetry while the real view preserves local flow patterns, contrasting the two yields positive and negative samples that are simultaneously diverse and coherent. The paper reports that this dual design outperforms both undirected contrastive methods and existing digraph contrastive baselines on all seven

What carries the argument

The central mechanism is the personalized magnetic Laplacian perturbation, L_r,Δq = D_s − A_s ⊙ exp(iΘ_r^{q*+Δq}), where the per-edge charge q*_uv is derived from each node's in/out-degree uncertainty via a tanh coefficient, and a probabilistic factor r stochastically reverses edge orientations. This matrix views are fed into a shared complex-valued GCN that updates real and imaginary feature parts. The real-domain counterpart is a second-order direction-aware path sampler (BFS/DFS biased random walks) fed through a GRU aggregator. The two representations are concatenated and projected, then contrasted with inter-view and intra-view InfoNCE losses. The magnetic Laplacian carries global rotat

Load-bearing premise

The load-bearing premise is that a single value of the perturbation parameter q can simultaneously describe how every eigenvalue of the magnetic Laplacian changes—an assumption made in the proof of Theorem 2 via a mean-value argument, which does not hold for vector-valued functions.

What would settle it

Take a small digraph whose magnetic Laplacian has eigenvalues with differing q-derivatives, such as two disjoint directed cycles of different lengths, compute H_VN at q and q+Δq, and check whether the claimed bound |ΔH_VN| ≤ (β²/√n) σ_λ ||ΔL||_F holds for some Δq; a violation would disprove Theorem 2.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Unsupervised digraph representation learning can match or beat supervised GNNs on node classification and link prediction, reducing reliance on labeled data in directed networks.
  • Personalized, per-node perturbation—rather than uniform augmentation—appears to be a transferable principle for contrastive learning on heterogeneous graph topologies.
  • The dual-spatial design (global complex-phase view plus local path view) offers a general template for other self-supervised tasks where semantics live at different structural scales.
  • The magnetic Von Neumann entropy analysis, if correct, would provide a principled stability guarantee for magnetic perturbations, explaining why the augmentation preserves spectral structure.
  • The path-based real-domain branch suggests that explicit directional sequence modeling is a key ingredient for digraph contrastive learning beyond simple edge dropping or feature masking.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to apply the same dual-spatial recipe to temporal networks or knowledge graphs, where edge direction encodes time or ontology; the paper's conclusion hints at this but does not test it.
  • The ablation shows that removing either branch hurts performance, but it does not isolate whether the GRU's sequential modeling, rather than the path sampling itself, is the decisive factor; a simpler aggregator may retain most of the benefit.
  • The proof of Theorem 2 relies on a vector-valued mean-value step that does not hold in general, so the claimed boundedness of entropy variation is not established; the observed stability may instead stem from the small, node-adaptive Δq rather than a universal spectral bound.
  • Because the entropy monotonicity theorem depends on eigenvalue simplicity almost everywhere, the theoretical guarantees could fail on digraphs with repeated eigenvalues, even if the empirical method remains robust.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes S2-DiGCL, a self-supervised contrastive learning framework for directed graphs. It combines two augmentation views: a complex-domain "personalized magnetic Laplacian perturbation" that adjusts edge phases via topology-dependent charges, and a real-domain path-based subgraph augmentation using biased BFS/DFS walks encoded by a GRU. The two views are contrasted with an InfoNCE objective. The manuscript reports node classification and link prediction experiments on seven directed graph datasets and claims state-of-the-art results, along with a theoretical analysis of magnetic von Neumann entropy and its response to charge perturbations.

Significance. If the empirical results are correctly reported, S2-DiGCL would be a useful contribution to directed graph contrastive learning: it is one of the few methods that jointly use real- and complex-domain augmentations, and the idea of personalizing magnetic Laplacian perturbations by local in/out-degree uncertainty is interesting. The paper includes ablations, sensitivity analysis, convergence experiments, and comparisons against several directed-graph baselines, which are valuable. However, the central SOTA claim is undercut by inconsistencies in the paper's own tables, and the theoretical stability argument contains a genuine proof error. The contribution is therefore plausible but not yet established in its current form.

major comments (3)
  1. [Abstract and Section V-B.1, Tables II--III] The headline claim of 'SOTA performance with 4.41% improvement in node classification and 4.34% in link prediction' is not derived anywhere—no baseline set or aggregation rule is specified, and recomputation from Tables II and III does not transparently reproduce these numbers. More seriously, the text in Section V-B.1 states S2-DiGCL 'consistently achieves the highest accuracy across all datasets' and 'ranks first on six out of seven benchmarks and secures the second-best result on WikiCS,' but Table II shows S2-DiGCL is first on WikiCS (78.42 vs. GraphCL's 78.37), and Table III (Direction setting, Citeseer) shows S2-DiGCL at 86.37±1.69, below RAW-GNN's 86.79±0.15. The empirical claim is the paper's primary contribution, so these discrepancies must be corrected and the reported percentages explicitly computed.
  2. [Section IV-B, Theorem 2 and Eq. (27)] The proof of Theorem 2 is invalid. After applying the scalar mean value theorem to H_VN, the proof asserts: 'By a mean-value argument on each λ_k(q), there exists the same ξ such that Δλ = λ̇(ξ)Δq.' This is false for vector-valued functions; no single ξ exists in general (e.g., f(t)=(t²,t³)). Consequently, the bound σ_λ̇^{(ξ)} ≤ (1/√n)‖λ̇^{(ξ)}‖₂ ≤ (1/√n)‖ΔL‖_F/|Δq| in Eq. (27) does not follow, and Theorem 2's bound on entropy variation is unproven. Since Section IV-C uses this theorem to claim that the magnetic perturbation is stable, this is a load-bearing gap that needs a corrected proof or a substantially weakened statement.
  3. [Section III-A, Eqs. (4)--(6)] The construction of the randomized magnetic Laplacian is only Hermitian if the random phase matrix Θ_r^{(q*)} is skew-symmetric, i.e., Θ_r^{(q*)}(u,v) = -Θ_r^{(q*)}(v,u). The paper defines Φ_r^{(q*)}(u,v) per ordered pair in Eq. (4) and then applies it element-wise in Eq. (5), but it never states that the Bernoulli draws for (u,v) and (v,u) are coupled. If they are drawn independently, L_r,Δq^{(q*)} in Eq. (6) is not Hermitian, and the eigen-decomposition and spectral analysis used in Section IV do not apply. The authors must explicitly specify the coupling (e.g., one draw per unordered edge) or otherwise prove Hermiticity.
minor comments (4)
  1. [Notation, Eq. (9) vs. Section III-A] The symbol q is used both for the magnetic charge parameter and for the Node2Vec-style return parameter in Eq. (9). This is confusing in a paper where both appear in the same framework; consider renaming one of them.
  2. [Section V-C.2] The ablation description says 'w/o perturb and uniform perturb settings correspond to fixing the personalization coefficient r in Eq. (4) to 0.0 and 1.0, respectively,' but r is a probability, not a coefficient; the interpretation of r=0 and r=1 as 'without perturbation' vs. 'uniform perturbation' should be clarified.
  3. [Table IV] The 'Link' columns are not labeled as existence or direction prediction. Given the two settings in Table III, specify which one is reported in the ablation.
  4. [Section V-B.1] The sentence 'S2-DiGCL attains comparable or even superior performance to several supervised GNNs, including RAW-GNN and MagNet' is imprecise: on Pubmed, S2-DiGCL (79.16±0.81) is below GAT (79.83±0.30). Please qualify such claims to match the tables.

Circularity Check

0 steps flagged

No circular derivation; self-citations to MAP/DiRW are non-load-bearing, and the abstract SOTA percentages plus the Theorem 2 mean-value step are correctness/reproducibility concerns, not circularity.

full rationale

The central derivation is self-contained. The method is built from explicit design choices: Eq. (2)-(6) define a personalized charge from degree-balance entropy and a probabilistic phase perturbation; Eq. (9)-(11) define Node2Vec-style biased walks and a GRU aggregator; Eq. (13) is the InfoNCE objective. None of these quantities is defined from the reported labels, test accuracies, or link-prediction metrics, and no parameter is fitted to the headline SOTA targets. The only self-citations used as ingredients are MAP [30] for the U_v uncertainty formula and DiRW [26] for walk-based motivation; these are building blocks, not uniqueness claims or forbidden-alternative arguments, so they do not raise the circularity score. The theoretical section proves properties of magnetic von Neumann entropy; it does not feed back into the empirical results. Theorem 1's algebra is valid as far as it goes, but Theorem 2 in Section IV-B has a real proof gap: 'By a mean-value argument on each λ_k(q), there exists the same ξ such that Δλ = λ̇(ξ)Δq' invokes a common-ξ mean-value theorem for the vector-valued function λ(q), which is false for general vector functions. This is an omitted-proof/correctness concern, not a circular reduction. Similarly, the abstract's 4.41%/4.34% improvements are not traceable from Tables II-III, and the text is internally inconsistent (Table II shows S2-DiGCL highest on WikiCS while the text says it is second-best; Table III Direction/Citeseer is below RAW-GNN's 86.79±1.25). These are external-validity and reproducibility problems, not by-construction circularity. The conclusion's acknowledged computational-cost limitation is also irrelevant to circularity. Therefore no load-bearing circular step can be exhibited; the only self-citations are minor and non-load-bearing, giving a score of 2 at most.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method relies on standard spectral graph theory and on assumptions about the effect of magnetic Laplacian perturbations. The free parameters are mostly standard hyperparameters, but several are dataset-tuned (path length) and many are unreported, which is a reproducibility concern. The most critical unstated assumption is the Hermiticity of the perturbed Laplacian, and the most clearly false premise is the common-ξ mean value step in Theorem 2.

free parameters (6)
  • q0 (base charge)
    Global base charge controlling phase magnitude; a hyperparameter in Eq. (3), not reported in the paper.
  • r (perturbation probability)
    Controls likelihood of maintaining original edge orientation in Eq. (4); varied in sensitivity analysis (0.125-0.875), default not specified.
  • Δq (charge perturbation)
    Perturbation term in Eq. (6), constrained to keep q*+Δq in [0,0.25]; value not specified.
  • Path length l = CoraML=4, Citeseer=2, Pubmed=5, WikiCS=5
    Per-dataset tuned; results in Fig. 4(b) show different optimal values. This is a dataset-specific fit.
  • Node2Vec p and q
    Walk-bias parameters in Eq. (9); values not reported in the paper.
  • InfoNCE temperature τ
    Temperature in Eq. (13), not reported.
axioms (5)
  • standard math The magnetic Laplacian L(q) is Hermitian for q in [0,1/2) and admits real eigenvalues and eigendecomposition.
    Invoked in Section IV-B before Theorem 1; standard property from MagNet [20].
  • ad hoc to paper The random perturbation of edge phases preserves Hermiticity of the magnetic Laplacian.
    Eq. (5) defines Θ_r with independent draws per direction; unless the draws are coupled, the matrix is not necessarily Hermitian, violating the theoretical framework. This is an unstated assumption.
  • standard math Eigenvalues of L(q) are simple almost everywhere so they are differentiable a.e.
    Used in Theorem 1 proof; a generic spectral assumption.
  • ad hoc to paper The mean value theorem applies component-wise with a common ξ for all eigenvalue derivatives.
    This is the false step in Theorem 2's proof: for vector-valued functions, a common ξ does not generally exist. The entire bounded-entropy-variation theorem rests on it.
  • domain assumption Graph homophily guides the BFS/DFS split: BFS captures homophilous local structure, DFS captures heterophilous long-range structure.
    Used to justify real-domain path augmentation in Section III-B; based on [31] but not verified on the actual datasets.

pith-pipeline@v1.3.0-alltime-deepseek · 3700 in / 3594 out tokens · 117543 ms · 2026-08-04T09:14:51.513034+00:00 · methodology

0 comments
read the original abstract

Graph Contrastive Learning (GCL) has emerged as a powerful tool for extracting consistent representations from graphs, independent of labeled information. However, existing methods predominantly focus on undirected graphs, disregarding the pivotal directional information that is fundamental and indispensable in real-world networks (e.g., social networks and recommendations).In this paper, we introduce S2-DiGCL, a novel framework that emphasizes spatial insights from complex and real domain perspectives for directed graph (digraph) contrastive learning. From the complex-domain perspective, S2-DiGCL introduces personalized perturbations into the magnetic Laplacian to adaptively modulate edge phases and directional semantics. From the real-domain perspective, it employs a path-based subgraph augmentation strategy to capture fine-grained local asymmetries and topological dependencies. By jointly leveraging these two complementary spatial views, S2-DiGCL constructs high-quality positive and negative samples, leading to more general and robust digraph contrastive learning. Extensive experiments on 7 real-world digraph datasets demonstrate the superiority of our approach, achieving SOTA performance with 4.41% improvement in node classification and 4.34% in link prediction under both supervised and unsupervised settings.

Figures

Figures reproduced from arXiv: 2510.16311 by Daohan Su, Guoren Wang, Rong-Hua Li, Xunkai Li, Yang Zhang, Zhengyu Wu.

Figure 1
Figure 1. Figure 1: Illustration of naive (Upper) vs. personalized (Lower) augmentations [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of S2-DiGCL, including a complex-domain global personalized magnetic Laplacian augmentation (Left), a real-domain local path [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization results on CoraML. TABLE IV ABLATION STUDY ON CORAML AND CITESEER. Variant CoraML Citeseer Node Link Node Link w/o complex 71.41±1.48 82.29±2.10 62.68±1.49 76.54±3.41 w/o real 73.18±1.92 85.51±1.70 64.84±1.48 82.56±2.46 w/o perturb 69.81±2.12 84.21±2.15 61.17±0.57 79.29±1.42 uniform perturb 70.68±2.98 82.48±1.57 61.98±1.47 80.46±2.87 w/o direction 71.50±1.87 83.47±1.23 62.04±0.74 76.76±2.45 w… view at source ↗
Figure 4
Figure 4. Figure 4: Performance with different parameters. D. Sensitivity Analysis To address Q3, we perform sensitivity experiments to exam￾ine the influence of key hyperparameters on the performance of S2-DiGCL in the node classification task. The analysis is conducted on four representative datasets: CoraML, Citeseer, Pubmed, and WikiCS. We focus on two critical parameters that directly affect the dual-domain learning proc… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ChemHyperMag: Physics-informed magnetic hypergraph learning improves molecular ADMET prediction

    cs.LG 2026-07 conditional novelty 6.0

    A magnetic hypergraph neural network with chemistry-inspired directional flow improves ADMET property prediction on several benchmarks.

Reference graph

Works this paper leans on

49 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Measurement and analysis of online social networks,

    A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattachar- jee, “Measurement and analysis of online social networks,” inProceed- ings of the 7th ACM SIGCOMM conference on Internet measurement, 2007, pp. 29–42

    2007

  2. [2]

    Structure and evolution of online social networks,

    R. Kumar, J. Novak, and A. Tomkins, “Structure and evolution of online social networks,” inProceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, 2006, pp. 611– 617

    2006

  3. [3]

    Complex network topology of transportation systems,

    J. Lin and Y . Ban, “Complex network topology of transportation systems,”Transport reviews, vol. 33, no. 6, pp. 658–685, 2013

    2013

  4. [4]

    Complex network analysis of public transportation networks: A comprehensive study,

    A. H ´aznagy, I. Fi, A. London, and T. Nemeth, “Complex network analysis of public transportation networks: A comprehensive study,” in2015 International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS). IEEE, 2015, pp. 371– 378

    2015

  5. [5]

    Graph neural networks in recommender systems: a survey,

    S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui, “Graph neural networks in recommender systems: a survey,”ACM Computing Surveys, vol. 55, no. 5, pp. 1–37, 2022

    2022

  6. [6]

    Dcl: Diversified graph recommendation with contrastive learning,

    D. Su, B. Fan, Z. Zhang, H. Fu, and Z. Qin, “Dcl: Diversified graph recommendation with contrastive learning,”IEEE Transactions on Com- putational Social Systems, 2024

    2024

  7. [7]

    Graph neural networks for materials science and chemistry,

    P. Reiser, M. Neubert, A. Eberhard, L. Torresi, C. Zhou, C. Shao, H. Metni, C. van Hoesel, H. Schopmans, T. Sommeret al., “Graph neural networks for materials science and chemistry,”Communications Materials, vol. 3, no. 1, p. 93, 2022

    2022

  8. [8]

    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

  9. [9]

    Deep Graph Infomax,

    P. Veli ˇckovi´c, W. Fedus, W. L. Hamilton, P. Li `o, Y . Bengio, and R. D. Hjelm, “Deep Graph Infomax,” inInternational Conference on Learning Representations, ICLR, 2019

    2019

  10. [10]

    Deep graph contrastive representation learning,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Deep graph contrastive representation learning,”arXiv preprint arXiv:2006.04131, 2020

    Pith/arXiv arXiv 2006

  11. [11]

    Finding hierarchy in directed online social networks,

    M. Gupte, P. Shankar, J. Li, S. Muthukrishnan, and L. Iftode, “Finding hierarchy in directed online social networks,” inProceedings of the 20th international conference on World wide web, 2011, pp. 557–566

    2011

  12. [12]

    Street network studies: from networks to models and their representations,

    S. Marshall, J. Gil, K. Kropf, M. Tomko, and L. Figueiredo, “Street network studies: from networks to models and their representations,” Networks and Spatial Economics, vol. 18, pp. 735–749, 2018

    2018

  13. [13]

    Connectivity in a citation network: The development of dna theory,

    N. P. Hummon and P. Dereian, “Connectivity in a citation network: The development of dna theory,”Social networks, vol. 11, no. 1, pp. 39–63, 1989

    1989

  14. [14]

    How citation distortions create unfounded authority: analysis of a citation network,

    S. A. Greenberg, “How citation distortions create unfounded authority: analysis of a citation network,”Bmj, vol. 339, 2009

    2009

  15. [15]

    Why using either ag- gregated features or adjacency lists in directed or undirected graph? empirical study and simple classification method,

    S. Maekawa, Y . Sasaki, and M. Onizuka, “Why using either ag- gregated features or adjacency lists in directed or undirected graph? empirical study and simple classification method,”arXiv preprint arXiv:2306.08274, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Edge directionality improves learning on heterophilic graphs,

    E. Rossi, B. Charpentier, F. D. Giovanni, F. Frasca, S. G ¨unnemann, and M. Bronstein, “Edge directionality improves learning on heterophilic graphs,”in Proceedings of The European Conference on Machine Learn- ing and Principles and Practice of Knowledge Discovery in Databases, ECML-PKDD Workshop, 2023

    2023

  17. [17]

    Breaking the entanglement of homophily and heterophily in semi-supervised node classification,

    H. Sun, X. Li, Z. Wu, D. Su, R.-H. Li, and G. Wang, “Breaking the entanglement of homophily and heterophily in semi-supervised node classification,”arXiv preprint arXiv:2312.04111, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Directed graph contrastive learning,

    Z. Tong, Y . Liang, H. Ding, Y . Dai, X. Li, and C. Wang, “Directed graph contrastive learning,”Advances in neural information processing systems, vol. 34, pp. 19 580–19 593, 2021

    2021

  19. [19]

    Universal graph contrastive learning with a novel laplacian perturbation,

    T. Ko, Y . Choi, and C.-K. Kim, “Universal graph contrastive learning with a novel laplacian perturbation,” inUncertainty in Artificial Intelli- gence. PMLR, 2023, pp. 1098–1108

    2023

  20. [20]

    Magnet: A neural network for directed graphs,

    X. Zhang, Y . He, N. Brugnone, M. Perlmutter, and M. Hirn, “Magnet: A neural network for directed graphs,”Advances in Neural Information Processing Systems, NeurIPS, 2021

    2021

  21. [21]

    Graph con- trastive learning with augmentations,

    Y . You, T. Chen, Y . Sui, T. Chen, Z. Wang, and Y . Shen, “Graph con- trastive learning with augmentations,”Advances in neural information processing systems, vol. 33, pp. 5812–5823, 2020

    2020

  22. [22]

    Simgrace: A simple framework for graph contrastive learning without data augmentation,

    J. Xia, L. Wu, J. Chen, B. Hu, and S. Z. Li, “Simgrace: A simple framework for graph contrastive learning without data augmentation,” in Proceedings of the ACM Web Conference 2022, 2022, pp. 1070–1079

    2022

  23. [23]

    Sub-graph contrast for scalable self-supervised graph representation learning,

    Y . Jiao, Y . Xiong, J. Zhang, Y . Zhang, T. Zhang, and Y . Zhu, “Sub-graph contrast for scalable self-supervised graph representation learning,” in 2020 IEEE international conference on data mining (ICDM). IEEE, 2020, pp. 222–231

    2020

  24. [24]

    Directed graph convolutional network,

    Z. Tong, Y . Liang, C. Sun, D. S. Rosenblum, and A. Lim, “Directed graph convolutional network,”arXiv preprint arXiv:2004.13970, 2020

    Pith/arXiv arXiv 2004

  25. [25]

    Digraph inception convolutional networks,

    Z. Tong, Y . Liang, C. Sun, X. Li, D. Rosenblum, and A. Lim, “Digraph inception convolutional networks,”Advances in Neural Information Processing Systems, NeurIPS, 2020

    2020

  26. [26]

    Dirw: Path-aware digraph learning for heterophily,

    D. Su, X. Li, Z. Li, Y . Liao, R.-H. Li, and G. Wang, “Dirw: Path-aware digraph learning for heterophily,”arXiv preprint arXiv:2410.10320, 2024

    arXiv 2024

  27. [27]

    Representation learning on graphs with jumping knowledge networks,

    K. Xu, C. Li, Y . Tian, T. Sonobe, K.-i. Kawarabayashi, and S. Jegelka, “Representation learning on graphs with jumping knowledge networks,” inInternational Conference on Machine Learning, ICML, 2018

    2018

  28. [28]

    Graph attention multi-layer perceptron,

    W. Zhang, Z. Yin, Z. Sheng, Y . Li, W. Ouyang, X. Li, Y . Tao, Z. Yang, and B. Cui, “Graph attention multi-layer perceptron,”Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD, 2022

    2022

  29. [29]

    Graph signal processing for directed graphs based on the hermitian laplacian,

    S. Furutani, T. Shibahara, M. Akiyama, K. Hato, and M. Aida, “Graph signal processing for directed graphs based on the hermitian laplacian,” inJoint European Conference on Machine Learning and Knowledge Discovery in Databases, ECML-PKDD. Springer, 2020

    2020

  30. [30]

    Toward effective digraph representation learning: A magnetic adaptive propagation based approach,

    X. Li, D. Su, Z. Wu, G. Zeng, H. Qin, R.-H. Li, and G. Wang, “Toward effective digraph representation learning: A magnetic adaptive propagation based approach,” inProceedings of the ACM on Web Conference 2025, 2025, pp. 2908–2923

    2025

  31. [31]

    Birds of a feather: Homophily in social networks,

    M. McPherson, L. Smith-Lovin, and J. M. Cook, “Birds of a feather: Homophily in social networks,”Annual Review of SocioLoGy, vol. 27, no. 1, pp. 415–444, 2001

    2001

  32. [32]

    Node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “Node2vec: Scalable feature learning for networks,” inProceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD, 2016

    2016

  33. [33]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation,

    K. Cho, B. Van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using rnn encoder-decoder for statistical machine translation,”arXiv preprint arXiv:1406.1078, 2014

    Pith/arXiv arXiv 2014

  34. [34]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  35. [35]

    Approximate von neumann entropy for directed graphs,

    C. Ye, R. C. Wilson, C. H. Comin, L. d. F. Costa, and E. R. Hancock, “Approximate von neumann entropy for directed graphs,”Physical Review E, vol. 89, no. 5, p. 052804, 2014

    2014

  36. [36]

    Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking,

    A. Bojchevski and S. G ¨unnemann, “Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking,” inICLR Workshop on Representation Learning on Graphs and Manifolds, 2018

    2018

  37. [37]

    Revisiting semi- supervised learning with graph embeddings,

    Z. Yang, W. W. Cohen, and R. Salakhutdinov, “Revisiting semi- supervised learning with graph embeddings,” inInternational Confer- ence on Machine Learning, ICML, 2016

    2016

  38. [38]

    Geom-gcn: Geometric graph convolutional networks,

    H. Pei, B. Wei, K. C.-C. Chang, Y . Lei, and B. Yang, “Geom-gcn: Geometric graph convolutional networks,” inInternational Conference on Learning Representations, ICLR, 2020

    2020

  39. [39]

    Wiki-cs: A wikipedia-based benchmark for graph neural networks,

    P. Mernyei and C. Cangea, “Wiki-cs: A wikipedia-based benchmark for graph neural networks,”arXiv preprint arXiv:2007.02901, 2020

    Pith/arXiv arXiv 2007

  40. [40]

    Multi-Scale attributed node embedding,

    B. Rozemberczki, C. Allen, and R. Sarkar, “Multi-Scale attributed node embedding,”Journal of Complex Networks, vol. 9, no. 2, 05 2021

    2021

  41. [41]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Rep- resentations, ICLR, 2017

    2017

  42. [42]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,” inInternational Conference on Learning Representations, ICLR, 2018

    2018

  43. [43]

    Raw-gnn: Random walk aggregation based graph neural network,

    D. Jin, R. Wang, M. Ge, D. He, X. Li, W. Lin, and W. Zhang, “Raw-gnn: Random walk aggregation based graph neural network,” inProceedings of the International Joint Conference on Artificial Intelligence, IJCAI, 2022

    2022

  44. [44]

    Graph contrastive learning with adaptive augmentation,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Graph contrastive learning with adaptive augmentation,” inProceedings of the web con- ference 2021, 2021, pp. 2069–2080

    2021

  45. [45]

    Negative sampling strategies for contrastive self-supervised learning of graph representa- tions,

    H. Hafidi, M. Ghogho, P. Ciblat, and A. Swami, “Negative sampling strategies for contrastive self-supervised learning of graph representa- tions,”Signal Processing, vol. 190, p. 108310, 2022

    2022

  46. [46]

    Lightdic: A simple yet effective approach for large-scale digraph representation learning,

    X. Li, M. Liao, Z. Wu, D. Su, W. Zhang, R.-H. Li, and G. Wang, “Lightdic: A simple yet effective approach for large-scale digraph representation learning,”arXiv preprint arXiv:2401.11772, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.”Journal of machine learning research, vol. 9, no. 11, 2008. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13 Daohan Suis currently working toward the master’s degree in the School of Computer Science & Tech- nology, Beijing Institute of Technology. He received his bachelor’s degre...

    2008

  48. [2022]

    He has publishe 5+ papers in top DB/DM/AI conferences such as VLDB, WWW, AAAI

    His research interest lies in general Data- centric ML and Graph-ML in complex scenarios (di- rected, signed, hypergraph, and distributed). He has publishe 5+ papers in top DB/DM/AI conferences such as VLDB, WWW, AAAI. Rong-Hua Lireceived his PhD degree from the De- partment of Systems Engineering and Engineering Management, The Chinese University of Hong...

    2013

  49. [2024]

    His research interests lie in the areas of graph neural networks and large language models. Yang Zhangreceived his Bachelor’s degree in Computer Science and Technology from Beijing Institute of Technology in 2024, and his Master’s degree in Information Engineering from The Chinese University of Hong Kong in 2025. His research interests focus on graph neur...

    2024