Pith. sign in

REVIEW 4 major objections 5 minor 89 references

Are Heterogeneous Graph Neural Networks Truly Effective for Node Classification? A Causal Perspective

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Heterogeneous graph neural networks gain their edge from heterogeneous information itself — through increased homophily and local–global distribution discrepancy — not from model architecture or complexity, according to this causal audit.

desk verdict Big, careful benchmark; the causal mediation analysis defines its treatment from the outcome it is trying to explain, so the headline causal claims do not survive. read the letter →

arxiv 2510.05750 v2 pith:B2GYSFTD submitted 2025-10-07 cs.LG cs.AI

classification cs.LGcs.AI
keywords heterogeneousgraphneuralnetworksnodeclassificationcausalinferencehomophilylocal–globaldistributiondiscrepancyRGCNbenchmarkmediationanalysiscounterfactual
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

Given that HGNNs are usually assumed to be effective, this paper asks whether they truly are, and which ingredient matters: the model design or the heterogeneous data itself. The authors claim that the entire performance advantage of HGNNs over homogeneous GNNs comes from heterogeneous information, which improves node classification by increasing homophily and the difference between local and global label distributions, making classes more separable. They support this with a systematic reproduction of 20 baselines on 21 datasets, showing that a hyperparameter-tuned simple RGCN matches or beats most published HGNN results, and with a causal analysis using factual, counterfactual, doubly robust, and sensitivity methods. If right, the conclusion refocuses the field: architecture-driven advances may be overrated, while structure-aware uses of heterogeneity—through homophily and distribution discrepancy—are what matter.

What carries the argument

Three elements carry the argument. (1) The treatment indicator T(v)=1[π̂_G(v)>π̂_Ḡ(v)] marks whether a node is more accurately predicted under the heterogeneous graph G than under its homogeneous projection Ḡ. (2) Structural indicator patterns P1–P3 compare homophily H and local–global distribution discrepancy D between G and Ḡ (e.g., P1: H_G < H_avg^G and D_G < D_min^G). (3) A battery of causal estimators—factual ATE with BH-corrected significance, counterfactual uplift with probability of necessity/sufficiency, a doubly robust ATE under a minimal sufficient adjustment set, cross-method consistency (PSM, IPW, TMLE), and E-value sensitivity—support the mediation claim. The machinery works

What would settle it

Run an intervention that changes relation types or edge types independently of label structure—for example, randomly permuting edge type labels in a heterogeneous graph while keeping topology and features fixed—and measure whether node classification accuracy and homophily/distribution-discrepancy shift together. If accuracy is unchanged despite the structural indicators moving, the mediation claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the effectiveness of heterogeneous graph neural networks (HGNNs) for node classification is driven by heterogeneous information itself, not by model architecture or complexity. After reproducing 20 HGNN baselines across 21 datasets under a unified protocol and hyperparameter-tuning the simplest relation-based model RGCN, the authors find that tuned RGCN matches or exceeds most published results, concluding architecture has no causal effect. Comparing tuned RGCN on heterogeneous graphs with GCN on homogeneous projections shows consistent gains (often 5–10 points), attributed to heterogeneity. A three-step causal analysis defines node-level treatment as receiv

Load-bearing premise

The load-bearing assumption is that 'treated by heterogeneous information' can be defined as a node being more accurately classified under the heterogeneous graph than under the homogeneous projection; because the treatment indicator is defined from the accuracy outcome it is meant to explain, the causal mediation claims rest on that equivalence being valid.

Editorial extensions

If this is right

  • The tuned simple RGCN outcome implies that many published HGNN accuracy gaps may be artifacts of under-tuned baselines; performance claims should be re-benchmarked with retuned simple models.
  • Heterogeneous information, not architecture, is the causal source of gain; collapsing edge/node types consistently drops accuracy, so retaining relational semantics is necessary for the benefit.
  • Homophily and local–global distribution discrepancy are identified as the structural channels; interventions that strengthen either (e.g., by construction or rewiring) should improve separability.
  • The node-level structural patterns occur for a majority of nodes (57.67% for P1), indicating the mechanism is widespread rather than edge-case.

Reading between the lines

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

  • The paper does not test it, but if the mechanism is homophily-plus-distribution-discrepancy, then rewiring a homogeneous graph to increase these two quantities—without changing the model—should recover part of the heterogeneous gain; this is a direct, falsifiable extension.
  • The same argument suggests a practical design rule for HGNNs: optimize data structure and relation selection rather than message-passing machinery, and report hyperparameter-retuned RGCN as a mandatory baseline.
  • The framework could be carried over to other graph tasks (link prediction, graph classification) by replacing node classification accuracy with the relevant outcome, providing a way to test whether the same structural mediators appear.
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 a three-step causal audit of heterogeneous graph neural networks (HGNNs). Step 1 reproduces 20 HGNN baselines across 21 datasets and retunes RGCN, concluding that model architecture and complexity have no causal effect on node classification performance. Step 2 compares GCN on homogeneous projections with RGCN on heterogeneous graphs, concluding that heterogeneous information has a positive causal effect. Step 3 proposes a causal mediation analysis in which a node-level treatment indicator T(v) is defined as whether the node is more accurately classified under heterogeneous information than under the homogeneous projection, and candidate mediators are structural indicators such as homophily and local–global distribution discrepancy. The paper concludes that heterogeneous information improves performance by increasing homophily and local–global distribution discrepancy, and that architecture is not the decisive factor.

Significance. If the conclusions held, the paper would provide an important and somewhat provocative reassessment of HGNN design, aligning with some prior empirical findings that simple baselines are competitive. The benchmark reproduction effort, including public code, is a useful community resource. However, the central causal claims are not supported by the analyses as presented. The Step 3 mediation framework defines its treatment from the outcome being explained, and the Step 1 and Step 2 comparisons conflate tuning budgets and architecture with information content. These issues are load-bearing for the paper's main conclusions, so the significance of the work depends on corrections that are not local.

major comments (4)
  1. [§4.3.2, Eq. (21), Eq. (23), Tables 4–6] The treatment indicator T(v) = 1[π_G(v) > π_{Ḡ}(v)] is defined directly from the outcome the paper seeks to explain: whether accuracy is higher under heterogeneous information. This is not an intervention or a manipulable treatment; it is the sign of the realized accuracy difference. Consequently, the potential-outcomes assumptions in §3.1 (SUTVA, ignorability, overlap) are incoherent for this 'treatment' — there is no counterfactual under a manipulation that sets T(v). The mediator indicators Z_{M,A}(v) in Eq. (23) are also defined from the same heterogeneous-vs-homogeneous contrast, making them post-treatment relative to T. Tables 4–6 therefore report associations between two outcome-derived variables, not causal effects. The E-value, MSAS, and cross-method consistency checks cannot repair this, because no adjustment on covariates converts an outcome-defined variable into a valid caus
  2. [§4.1, Table 2] The claim that 'model architecture and complexity have no causal effect on node classification performance' is not supported by the experimental design in Step 1. The comparison in Table 2 is between RGCN with comprehensive hyperparameter retuning performed by the authors and each baseline reproduced with its paper-reported hyperparameters. These are not matched tuning budgets. The observed RGCN gains could reflect the additional tuning effort rather than the absence of an architectural effect. To draw the stated conclusion, all baselines would need to be tuned under comparable procedures and budgets, or the analysis would need to explicitly vary architecture while controlling for tuning.
  3. [§4.2, Table 3] Step 2 compares GCN on homogeneous projections with RGCN on heterogeneous graphs. This changes two things simultaneously: the input graph and the model architecture. RGCN has relation-specific parameters and a different message-passing mechanism; GCN does not. Thus the performance gap in Table 3 cannot be unambiguously attributed to heterogeneous information. The text asserts that 'any performance differences can be attributed to heterogeneous information, rather than to architectural design,' but this requires that the two architectures are matched in capacity and inductive bias, which is not established. A proper control would use the same model class on both graphs (e.g., RGCN applied to the homogeneous projection, or GCN augmented to consume edge types) with matched hyperparameters.
  4. [§3.2, Tables 2–3] The benchmark comparisons are reported without error bars or significance tests. Claims such as 'RGCN consistently outperforms GCN across datasets' are based on point estimates from presumably few runs; some differences are small (e.g., DBLP Macro F1 94.65 vs 94.86 for SlotGAT in Table 2). Given the strong causal conclusions drawn, the lack of repeated-run variability and statistical testing (including correction for multiple comparisons, even though the paper describes such methodology in §3.2) is a serious gap. This is compounded by the unequal-tuning issue in Step 1.
minor comments (5)
  1. [§4.3.3, Table 4] The criteria for 'PASS' in Table 4 are not specified. The paper mentions significance and effect-size criteria but does not state the threshold for ATE, SE, or RR, nor the significance level used. Please provide concrete decision rules.
  2. [§4.3.6, Table 6] The 'Consistency' column reports 5/5 for each pattern, but the actual estimates from the five estimators are not shown. To assess cross-method consistency, the reader needs the point estimates and confidence intervals for each method, not just a summary count.
  3. [§4.3.5] The MSAS definition in Eq. (25) is abstract and does not specify which covariates Z* are used in the analysis. 'Adjusting for H' and 'adjusting for D' are mentioned, but the actual covariate set and the way it is constructed from the structural indicators are unclear.
  4. [§4.3.4] The counterfactual analysis in Table 5 uses simulated node-level outcomes, but the simulation procedure is not described. PN and PS are defined only verbally; the formulas or algorithmic steps for computing them should be provided.
  5. [Table 2] Several cells are missing (e.g., RGCN* values for many baselines) and some entries are formatted inconsistently. A complete table with all reported and reproduced values, or an explicit note on why some values are absent, would improve readability.

Circularity Check

1 steps flagged · score 8.0 of 10

Step 3's causal mediation analysis is circular: Eq. (21) defines the 'treatment' as the outcome contrast itself, and Eq. (23) defines the mediators from the same contrast, so the claimed homophily/distribution-discrepancy effects are associations among outcome-derived variables.

  1. self definitional [Section 4.3.2, Eqs. (21)–(23); Tables 4–6; conclusion]
    "We then define a treatment indicator T(v) = 1[π̂_G(v) > π̂_{̅G}(v)], which equals 1 if node v achieves higher predictive accuracy under heterogeneous information. ... Together, the treatment variable T(v) and the factor indicators {Z_{M,A}(v)} provide node-level factual evidence on how heterogeneous information affects predictive performance and structural properties."

    The paper's own framework says the treatment should be 'the introduction of heterogeneous relation information' and the outcome should be classification performance. Instead, Eq. (21) defines the treatment T(v) as the indicator that heterogeneous information produced higher accuracy for node v—i.e., the very outcome contrast the paper is trying to explain. Eq. (23) defines the mediator indicators Z from the same heterogeneous-versus-homogeneous graph comparison. Tables 4–6 then report ATEs, risk ratios, uplifts, E-values, and doubly robust estimates relating these two outcome-derived variables. Since T has no independent manipulation and Z is measured post-contrast, the 'positive causal effect through homophily and local–global distribution discrepancy' reduces by construction to an associ

full rationale

The benchmark and reproduction components (Steps 1 and 2) are self-contained empirical exercises: tuned RGCN is compared with official baselines and GCN-on-homogeneous versus RGCN-on-heterogeneous accuracies are reported; those comparisons are not circular, though Step 2's architecture/heterogeneity confound is a validity concern rather than a definitional circle. The load-bearing causal claim in the abstract and conclusion, however, rests on Step 3's mediation analysis, and that analysis is circular. Eq. (21) defines the binary 'treatment' T(v) as whether RGCN on the heterogeneous graph beats GCN on the homogeneous projection for node v—i.e., as the very performance contrast the paper seeks to attribute to structural factors. Eq. (23) defines the candidate mediators as indicators that the same contrast moves homophily or local–global distribution discrepancy in the hypothesized direction. Tables 4–6 then present ATEs, counterfactual uplifts, consistency checks, and E-values relating these two outcome-derived variables. Because T has no independent counterfactual and Z is measured from the same contrast, no propensity adjustment, TMLE, or sensitivity analysis can convert the reported associations into causal effects of homophily and distribution discrepancy on classification. The central scientific conclusion therefore reduces by definition to the outcome indicator in Eq. (21). Score 8: the claimed causal mediation result is forced by construction, although the paper's non-causal benchmark evidence retains independent content.

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

The causal claims require strong assumptions that are not met: outcome-derived treatment, mediator screening on the outcome, and unequal tuning. No new physical entities are introduced.

free parameters (2)
  • RGCN per-dataset hyperparameters = not reported in paper
    Step 1's 'architecture has no causal effect' relies on RGCN being tuned at least as thoroughly as baselines; the search space/budget is not described.
  • Structural indicator patterns P1/P2/P3 = selected post hoc from candidate indicators
    The abstract says mediators were screened by 'treatment-induced changes and their associations with performance improvement'; only three patterns are retained, with no full multiple-testing audit.
assumptions (4)
  • domain assumption SUTVA, ignorability, and overlap hold for the node-level treatment contrast.
    Invoked in Section 3.1, but node outcomes interfere through message passing and the treatment is outcome-defined, so these assumptions are not justified.
  • domain assumption Homophily and local-global distribution discrepancy are the relevant mediating structural factors.
    Section 4.3.1 selects these two factors by citing prior work; the causal mediation claim depends on this selection being complete.
  • ad hoc to paper RGCN and GCN are canonical representatives such that comparing them isolates the effect of heterogeneous information.
    Section 4.2 uses tuned RGCN on heterogeneous graphs versus GCN on homogeneous projections; architecture and hyperparameters differ, so the contrast is not solely heterogeneity.
  • ad hoc to paper Screening mediators by their association with the outcome does not invalidate downstream causal inference.
    The abstract and Section 4.3.2 describe screening by treatment-induced changes and association with performance; this is post-hoc selection that affects the causal estimates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are Heterogeneous Graph Neural Networks Truly Effective for Node Classification? A Causal Perspective." pith.science (2026). https://pith.science/paper/B2GYSFTD

@misc{pith2026251005750,
  author       = {Pith},
  title        = {Pith review of: Are Heterogeneous Graph Neural Networks Truly Effective for Node Classification? A Causal Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2GYSFTD}},
  note         = {Machine review of arXiv:2510.05750}
}
read the original abstract

Graph neural networks (GNNs) have achieved remarkable success in node classification. Building on this progress, heterogeneous graph neural networks (HGNNs) integrate relation types and node and edge semantics to leverage heterogeneous information. Causal analysis for HGNNs is advancing rapidly, aiming to separate genuine causal effects from spurious correlations. However, whether HGNNs are intrinsically effective for node classification remains underexamined, and most studies implicitly assume rather than establish this effectiveness. In this work, we examine HGNNs for node classification from two perspectives: model architecture and heterogeneous information. We conduct a systematic reproduction across 21 datasets and 20 baselines, complemented by comprehensive hyperparameter retuning. To further disentangle the source of performance gains, we develop a causal mediation analysis framework that treats the introduction of heterogeneous relation information as the treatment, candidate structural properties as mediators, and node classification performance as the outcome. This framework first screens candidate mediators according to their treatment-induced changes and their associations with performance improvement, and then decomposes the total effect into mediated and direct effects. Our results lead to two conclusions. First, model architecture and complexity have no causal effect on node classification performance. Second, heterogeneous information exerts a positive causal effect primarily through increasing homophily and local-global distribution discrepancy, which makes node classes more distinguishable. The implementation is publicly available at https://github.com/YXNTU/CausalHGNN.

Figures

Figures reproduced from arXiv: 2510.05750 by the authors.

Figure 1
Figure 1. Overview of our research roadmap for disentangling the causal effects of hetero [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

89 extracted references · 11 linked inside Pith

  1. [1]

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, S. Y. Philip, A compre- hensive survey on graph neural networks, IEEE transactions on neural networks and learning systems 32 (1) (2020) 4–24

  2. [2]

    F. Xia, K. Sun, S. Yu, A. Aziz, L. Wan, S. Pan, H. Liu, Graph learning: A survey, IEEE Transactions on Artificial Intelligence 2 (2) (2021) 109– 127

  3. [3]

    Y. Bei, W. Chen, H. Chen, S. Zhou, C. Yang, J. Fan, L. Huang, J. Bu, Correlation-aware graph convolutional networks for multi-label node classification, in: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, 2025, pp. 37–48

  4. [4]

    X. Li, J. Saúde, Explain graph neural networks to understand weighted graph features in node classification, in: International Cross-Domain Conference for Machine Learning and Knowledge Extraction, Springer, 2020, pp. 57–76

  5. [5]

    S. Gong, J. Ni, N. Sachdeva, C. Yang, W. Jin, Gc4nc: A benchmark framework for graph condensation on node classification with new in- sights, arXiv preprint arXiv:2406.16715 (2024)

  6. [6]

    X. Zhao, H. Chen, Z. Xing, C. Miao, Brain-inspired search engine assis- tant based on knowledge graph, IEEE Transactions on Neural Networks and Learning Systems 34 (8) (2021) 4386–4400

  7. [7]

    A. Ali, J. Li, H. Chen, A. B. A. Ajlouni, A binary classification social network dataset for graph machine learning, arXiv preprint arXiv:2503.02397 (2025). 22

  8. [8]

    Ma, Z.-D

    Z.-X. Ma, Z.-D. Chen, L.-J. Zhao, Z.-C. Zhang, X. Luo, X.-S. Xu, Cross- layer and cross-sample feature optimization network for few-shot fine- grained image classification, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 38, 2024, pp. 4136–4144

Show all 89 references
  1. [9]

    Sharma, Y.-C

    K. Sharma, Y.-C. Lee, S. Nambi, A. Salian, S. Shah, S.-W. Kim, S. Ku- mar, A survey of graph neural networks for social recommender systems, ACM Computing Surveys 56 (10) (2024) 1–34

  2. [10]

    H. Li, S. Li, J. Sun, Z. Xing, X. Peng, M. Liu, X. Zhao, Improving api caveats accessibility by mining api caveats knowledge graph, in: 2018 IEEE International Conference on Software Maintenance and Evolution (ICSME), IEEE, 2018, pp. 183–193

  3. [11]

    X. Yang, X. Zhao, Z. Shen, A generalizable anomaly detection method in dynamic graphs, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, pp. 22001–22009

  4. [12]

    H. Du, L. Yuan, Q. Yang, X. Chen, Y. Zhao, H. Ji, F. Zhuang, C. Yang, G. Kou, Identifying evidence subgraphs for financial risk de- tection via graph counterfactual and factual reasoning, arXiv preprint arXiv:2503.06441 (2025)

  5. [13]

    X. Zhao, S. Liu, S.-Y. Yang, C. Miao, Medrag: Enhancing retrieval- augmented generation with knowledge graph-elicited reasoning for healthcare copilot, in: Proceedings of the ACM on Web Conference 2025, 2025, pp. 4442–4457

  6. [14]

    X. Zhao, S. Liu, S.-Y. Yang, C. Miao, A smart multimodal healthcare copilot with powerful llm reasoning, arXiv preprint arXiv:2506.02470 (2025)

  7. [15]

    X. Li, N. C. Dvornek, Y. Zhou, J. Zhuang, P. Ventola, J. S. Duncan, Graph neural network for interpreting task-fmri biomarkers, in: Inter- national conference on medical image computing and computer-assisted intervention, Springer, 2019, pp. 485–493

  8. [16]

    J. Qiao, W. Gao, J. Jin, D. Wang, X. Guo, B. Manavalan, L. Wei, Molecular pretraining models towards molecular property prediction, Science China Information Sciences 68 (7) (2025) 170104. 23

  9. [17]

    J. Qiao, J. Jin, D. Wang, S. Teng, J. Zhang, X. Yang, Y. Liu, Y. Wang, L. Cui, Q. Zou, et al., A self-conformation-aware pre-training framework for molecular property prediction with substructure interpretability, Na- ture Communications 16 (1) (2025) 4382

  10. [18]

    Velickovic, G

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y. Bengio, et al., Graph attention networks, stat 1050 (20) (2017) 10–48550

  11. [19]

    Hamilton, Z

    W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large graphs, Advances in neural information processing systems 30 (2017)

  12. [20]

    Zhao, Explainable q&a system based on domain-specific knowledge graph (2021)

    X. Zhao, Explainable q&a system based on domain-specific knowledge graph (2021)

  13. [21]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl, Neural message passing for quantum chemistry, in: International conference on machine learning, PMLR, 2017, pp. 1263–1272

  14. [22]

    X. Yang, X. Zhao, Z. Shen, A unified gradient regularization method for heterogeneous graph neural networks, Neural Networks (2025) 108104

  15. [23]

    C. Shi, Y. Li, J. Zhang, Y. Sun, S. Y. Philip, A survey of heterogeneous information network analysis, IEEE Transactions on Knowledge and Data Engineering 29 (1) (2016) 17–37

  16. [24]

    C. Yang, Y. Xiao, Y. Zhang, Y. Sun, J. Han, Heterogeneous network representation learning: A unified framework with survey and bench- mark, IEEE Transactions on Knowledge and Data Engineering 34 (10) (2020) 4854–4873

  17. [25]

    Zhang, D

    C. Zhang, D. Song, C. Huang, A. Swami, N. V. Chawla, Heterogeneous graph neural network, in: Proceedings of the 25th ACM SIGKDD in- ternational conference on knowledge discovery & data mining, 2019, pp. 793–803

  18. [26]

    X. Wang, D. Bo, C. Shi, S. Fan, Y. Ye, S. Y. Philip, A survey on heterogeneous graph embedding: methods, techniques, applications and sources, IEEE Transactions on Big Data 9 (2) (2022) 415–436. 24

  19. [27]

    Y. Yang, J. Chen, X. Gao, Y. Xiang, Dual de-confounded causal inter- vention method for knowledge graph error detection, Knowledge-Based Systems 305 (2024) 112644

  20. [28]

    T. Ban, L. Chen, D. Lyu, X. Wang, Q. Zhu, H. Chen, Llm-driven causal discovery via harmonized prior, IEEE Transactions on Knowledge and Data Engineering (2025)

  21. [29]

    Y.Chen, J.Cao, Y.Wang, J.Wu, H.Chen, G.Xu, Causalvariationalin- ference for deconfounded multi-behavior recommendation, ACM Trans- actions on Information Systems (2025)

  22. [30]

    J. Zhu, X. Wu, M. Usman, X. Wang, H. Chen, Link prediction in continuous-time dynamic heterogeneous graphs with causality of event types, International Journal of Crowd Science 6 (2) (2022) 80–91

  23. [31]

    B. Jing, D. Zhou, K. Ren, C. Yang, Causality-aware spatiotemporal graph neural networks for spatiotemporal time series imputation, in: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 1027–1037

  24. [32]

    Adhikari, E

    S. Adhikari, E. Zheleva, Inferring individual direct causal effects under heterogeneous peer influence, Machine Learning 114 (4) (2025) 113

  25. [33]

    Z. Zhao, Y. Bai, R. Xiong, Q. Cao, C. Ma, N. Jiang, F. Wu, K. Kuang, Learning individual treatment effects under heterogeneous interference in networks, ACM Transactions on Knowledge Discovery from Data 18 (8) (2024) 1–21

  26. [34]

    Y. Wu, Y. Yuan, Estimating heterogeneous causal effect on networks via orthogonal learning, arXiv preprint arXiv:2509.18484 (2025)

  27. [35]

    Q. Lv, M. Ding, Q. Liu, Y. Chen, W. Feng, S. He, C. Zhou, J. Jiang, Y. Dong, J. Tang, Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks, in: Proceedings of the 27th ACM SIGKDD conference on knowledge dis- covery & dat...

  28. [36]

    Schlichtkrull, T

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, M.Welling, Modelingrelationaldatawithgraphconvolutionalnetworks, 25 in: The semantic web: 15th international conference, ESWC 2018, Her- aklion, Crete, Greece, June 3–7, 2018, proceedings 15, Springer, 2018, pp...

  29. [37]

    Y. Ma, X. Liu, N. Shah, J. Tang, Is homophily a necessity for graph neural networks?, arXiv preprint arXiv:2106.06134 (2021)

  30. [38]

    J. Zhu, Y. Yan, L. Zhao, M. Heimann, L. Akoglu, D. Koutra, Beyond homophily in graph neural networks: Current limitations and effective designs, Advances in neural information processing systems 33 (2020) 7793–7804

  31. [39]

    S. Luan, C. Hua, M. Xu, Q. Lu, J. Zhu, X.-W. Chang, J. Fu, J. Leskovec, D. Precup, When do graph neural networks help with node classifica- tion? investigating the homophily principle on node distinguishability, Advances in Neural Information Processing Systems 36 (2024)

  32. [40]

    J. Zhu, G. Li, Y.-A. Yang, J. Zhu, X. Cui, D. Koutra, On the impact of feature heterophily on link prediction with graph neural networks, Advances in Neural Information Processing Systems 37 (2024) 65823– 65851

  33. [41]

    Z. Liu, Y. Fang, C. Liu, S. C. Hoi, Node-wise localization of graph neural networks, arXiv preprint arXiv:2110.14322 (2021)

  34. [42]

    Yehudai, E

    G. Yehudai, E. Fetaya, E. Meirom, G. Chechik, H. Maron, From local structures to size generalization in graph neural networks, in: Interna- tional Conference on Machine Learning, PMLR, 2021, pp. 11975–11986

  35. [43]

    Loveland, J

    D. Loveland, J. Zhu, M. Heimann, B. Fish, M. T. Schaub, D. Koutra, On performance discrepancies across local homophily levels in graph neural networks, in: Learning on Graphs Conference, PMLR, 2024, pp. 6–1

  36. [44]

    Z. Hu, Y. Dong, K. Wang, Y. Sun, Heterogeneous graph transformer, in: Proceedings of the web conference 2020, 2020, pp. 2704–2710

  37. [45]

    X. Wang, H. Ji, C. Shi, B. Wang, Y. Ye, P. Cui, P. S. Yu, Heterogeneous graph attention network, in: The world wide web conference, 2019, pp. 2022–2032. 26

  38. [46]

    X. Fu, J. Zhang, Z. Meng, I. King, Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding, in: Proceedings of the web conference 2020, 2020, pp. 2331–2341

  39. [47]

    S. Yun, M. Jeong, R. Kim, J. Kang, H. J. Kim, Graph transformer networks, Advances in neural information processing systems 32 (2019)

  40. [48]

    J. Zhao, X. Wang, C. Shi, Z. Liu, Y. Ye, Network schema preserving heterogeneous information network embedding, in: International joint conference on artificial intelligence (IJCAI), 2020

  41. [49]

    Q. Mao, Z. Liu, C. Liu, J. Sun, Hinormer: Representation learning on heterogeneous information networks with graph transformer, in: Pro- ceedings of the ACM Web Conference 2023, 2023, pp. 599–610

  42. [50]

    X. Yang, M. Yan, S. Pan, X. Ye, D. Fan, Simple and efficient hetero- geneous graph neural network, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 37, 2023, pp. 10816–10824

  43. [51]

    D. B. Rubin, Estimating causal effects of treatments in randomized and nonrandomizedstudies., JournalofeducationalPsychology66(5)(1974) 688

  44. [52]

    Sekhon, The neyman—rubin model of causal inference and estimation via matching methods (2008)

    J. Sekhon, The neyman—rubin model of causal inference and estimation via matching methods (2008)

  45. [53]

    Cotta, B

    L. Cotta, B. Bevilacqua, N. Ahmed, B. Ribeiro, Causal lifting and link prediction, Proceedings of the Royal Society A 479 (2276) (2023) 20230121

  46. [54]

    G. Chen, Y. Wang, F. Guo, Q. Guo, J. Shao, H. Shen, X. Cheng, Causal- ity and independence enhancement for biased node classification, in: Proceedings of the 32nd ACM international conference on information and knowledge management, 2023, pp. 203–212

  47. [55]

    J. Sun, Y. Sheng, L. He, Y. Qin, M. Liu, T. Jia, Cegrl-tkgr: A causal enhanced graph representation learning framework for temporal knowl- edge graph reasoning, arXiv preprint arXiv:2408.07911 (2024). 27

  48. [56]

    Jiang, Y

    S. Jiang, Y. Sun, Estimating causal effects on networked observational data via representation learning, in: Proceedings of the 31st ACM Inter- national Conference on Information & Knowledge Management, 2022, pp. 852–861

  49. [57]

    X. Du, J. Li, D. Cheng, L. Liu, W. Gao, X. Chen, Estimating peer direct and indirect effects in observational network data, arXiv preprint arXiv:2408.11492 (2024)

  50. [58]

    Farzam, A

    A. Farzam, A. R. Tannenbaum, G. Sapiro, From geometry to causality- ricci curvature and the reliability of causal inference on networks, in: Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024, OpenReview.net, 2024. URLhttps:/...

  51. [59]

    R. Cai, Z. Yang, W. Chen, Y. Yan, Z. Hao, Generalization bound for es- timating causal effects from observational network data, in: Proceedings of the 32nd ACM International Conference on Information and Knowl- edge Management, 2023, pp. 163–172

  52. [60]

    B. Hu, Z. An, Z. Wu, K. Tu, Z. Liu, Z. Zhang, J. Zhou, Y. Feng, J. Chen, Graph disentangle causal model: Enhancing causal inference in networked observational data, in: Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining, 2025, pp. 1–9

  53. [61]

    Y. Sui, C. Tang, Z. Chu, J. Fang, Y. Gao, Q. Cui, L. Li, J. Zhou, X. Wang, Invariant graph learning for causal effect estimation, in: Pro- ceedings of the ACM Web Conference 2024, 2024, pp. 2552–2562

  54. [62]

    X. Wang, T. Ban, L. Chen, D. Lyu, Q. Zhu, H. Chen, Large-scale hier- archical causal discovery via weak prior knowledge, IEEE Transactions on Knowledge and Data Engineering (2025)

  55. [63]

    T. Ban, X. Wang, L. Chen, D. Lyu, X. Fan, H. Chen, Harnessing the power of knowledge graphs to improve causal discovery, IEEE Transac- tions on Emerging Topics in Computational Intelligence (2025)

  56. [64]

    M. B. Begum, A. Yogeshwaran, N. Nagarajan, P. Rajalakshmi, Dynamic network security leveraging efficient covinet with granger causality- 28 inspired graph neural networks for data compression in cloud iot devices, Knowledge-Based Systems 309 (2025) 112859

  57. [65]

    H. Gao, P. Qiao, Y. Jin, F. Wu, J. Li, C. Zheng, Introducing diminutive causal structure into graph representation learning, Knowledge-Based Systems 293 (2024) 111592

  58. [66]

    Khaled, A

    A. Khaled, A. M. T. Elsir, P. Wang, Y. Shen, Q. Zhang, A graph- based approach for traffic prediction using similarity and causal relations between nodes, Knowledge-Based Systems 296 (2024) 111913

  59. [67]

    W. G. Cochran, Theχ2 test of goodness of fit, The Annals of mathe- matical statistics (1952) 315–345

  60. [68]

    Agresti, M

    A. Agresti, M. Kateri, Categorical data analysis, in: International en- cyclopedia of statistical science, Springer, 2025, pp. 408–411

  61. [69]

    Benjamini, Y

    Y. Benjamini, Y. Hochberg, Controlling the false discovery rate: a prac- tical and powerful approach to multiple testing, Journal of the Royal statistical society: series B (Methodological) 57 (1) (1995) 289–300

  62. [70]

    Jaeschke, J

    R. Jaeschke, J. Singer, G. H. Guyatt, Measurement of health status: as- certaining the minimal clinically important difference, Controlled clini- cal trials 10 (4) (1989) 407–415

  63. [71]

    H. Han, T. Zhao, C. Yang, H. Zhang, Y. Liu, X. Wang, C. Shi, Open- hgnn: anopen sourcetoolkitforheterogeneousgraphneuralnetwork, in: Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp. 3993–3997

  64. [72]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph con- volutional networks, arXiv preprint arXiv:1609.02907 (2016)

  65. [73]

    Y. Luo, L. Shi, X.-M. Wu, Classic gnns are strong baselines: Reassessing gnns for node classification, arXiv preprint arXiv:2406.08993 (2024)

  66. [74]

    S. Zhu, C. Zhou, S. Pan, X. Zhu, B. Wang, Relation structure-aware heterogeneous graph neural network, in: 2019 IEEE international con- ference on data mining (ICDM), IEEE, 2019, pp. 1534–1539. 29

  67. [75]

    P. Yu, C. Fu, Y. Yu, C. Huang, Z. Zhao, J. Dong, Multiplex heteroge- neous graph convolutional network, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 2377–2387

  68. [76]

    Z. Zhou, J. Shi, R. Yang, Y. Zou, Q. Li, Slotgat: slot-based message passing for heterogeneous graphs, in: International Conference on Ma- chine Learning, PMLR, 2023, pp. 42644–42657

  69. [77]

    J. Hu, B. Hooi, B. He, Efficient heterogeneous graph learning via ran- domprojection, IEEETransactionsonKnowledgeandDataEngineering (2024)

  70. [78]

    Y. Mo, F. Nie, P. Hu, H. T. Shen, Z. Zhang, X. Wang, X. Zhu, Self- supervised heterogeneous graph learning: a homophily and heterogene- ity view, in: The Twelfth International Conference on Learning Repre- sentations

  71. [79]

    X. Zhao, Y. Liu, Y. Xu, Y. Yang, X. Luo, C. Miao, Heterogeneous star graph attention network for product attributes prediction, Advanced Engineering Informatics 51 (2022) 101447

  72. [80]

    34, 2020, pp

    H.Hong, H.Guo, Y.Lin, X.Yang, Z.Li, J.Ye, Anattention-basedgraph neural network for heterogeneous structural learning, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 34, 2020, pp. 4132– 4139

  73. [81]

    X. Wang, N. Liu, H. Han, C. Shi, Self-supervised heterogeneous graph neural network with co-contrastive learning, in: Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 1726–1736

  74. [82]

    X. Li, D. Ding, B. Kao, Y. Sun, N. Mamoulis, Leveraging meta-path contexts for classification in heterogeneous information networks, in: 2021 IEEE 37th International Conference on Data Engineering (ICDE), IEEE, 2021, pp. 912–923

  75. [83]

    J. Zhao, X. Wang, C. Shi, B. Hu, G. Song, Y. Ye, Heterogeneous graph structure learning for graph neural networks, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 35, 2021, pp. 4697–4705. 30

  76. [84]

    J. Guo, L. Du, W. Bi, Q. Fu, X. Ma, X. Chen, S. Han, D. Zhang, Y. Zhang, Homophily-oriented heterogeneous graph rewiring, in: Pro- ceedings of the ACM Web Conference 2023, 2023, pp. 511–522

  77. [85]

    X. Fu, I. King, Mecch: metapath context convolution-based heteroge- neous graph neural networks, Neural Networks 170 (2024) 266–275

  78. [86]

    Z. Lu, Y. Fang, C. Yang, C. Shi, Heterogeneous graph transformer with poly-tokenization, International Joint Conferences on Artificial Intelli- gence, 2024

  79. [87]

    C.Li, J.Fu, Y.Yan, Z.Zhao, Q.Zeng, Higherorderheterogeneousgraph neural network based on node attribute enhancement, Expert Systems with Applications 238 (2024) 122404

  80. [88]

    C. Park, D. Kim, J. Han, H. Yu, Unsupervised attributed multiplex network embedding, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 34, 2020, pp. 5371–5378

  81. [89]

    H. Pei, B. Wei, K. C.-C. Chang, Y. Lei, B. Yang, Geom-gcn: Geometric graph convolutional networks, arXiv preprint arXiv:2002.05287 (2020). 31

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.