Pith. sign in

REVIEW 4 major objections 5 minor 30 references

THGFM: Dual-Branch Temporal Heterogeneous Graph Fusion Model

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read THGFM couples shared-space and relation-specific attention branches, fuses them with type-conditioned non-competitive gates, and injects relative time into attention scoring, beating the standard heterogeneous graph transformer baseline on

desk verdict THGFM is a careful, well-disclosed dual-branch architecture with disciplined evaluation, but its headline +3.25% gain over HGT rests on an unanchored reimplementation — treat the numbers as plausible, not established. read the letter →

arxiv 2607.27303 v1 pith:YI5HGQEB submitted 2026-07-29 cs.LG cs.CLstat.ML

classification cs.LGcs.CLstat.ML
keywords temporalheterogeneousgraphsgraphtransformersdual-branchattentionfusionrotarytype-conditionedgatedacademicbenchmarkslearningrelativetimeencoding
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 a temporal heterogeneous graph transformer can be improved by attacking two weaknesses at once: forcing a choice between shared cross-type parameters and relation-specific parameters, and treating time as an additive side feature rather than part of the attention decision. Its proposal, THGFM, runs two attention branches in parallel — one that projects all node types into a shared space, one that keeps type- and relation-specific projections — and merges them with type-conditioned non-competitive gated summation, so both branches can contribute at full strength when useful. It also introduces rotary temporal attention, which rotates queries and keys by half-phases of the relative time interval before scoring, making time directly reshape attention weights. The authors report that this recipe beats a reimplemented HGT baseline on every one of six tasks across four academic graph benchmarks, averaging a +3.25% relative gain with a peak of +12.37%, at comparable parameter counts.

What carries the argument

The central mechanism is the dual-path fusion module DRSF with its TC-NGSF operator: two independent feature-wise sigmoid gates, computed from the concatenated branch outputs (and a learnable target-type embedding in the type-conditioned variant), recalibrate the shared-space (SSTA) and relation-partitioned (RTTA) branch outputs before they are summed, so the two branches can be amplified or suppressed independently rather than forced into a zero-sum trade-off. The second supporting mechanism is RoTA (Rotary Temporal Attention), which rotates queries and keys by opposite half-phases proportional to relative time before the attention dot product, injecting time directly into score computation

What would settle it

Reproduce the baseline on OAG-CS with the original released code, preprocessing, and hyperparameters; if it yields a PF-L1 NDCG statistically above the paper's reported 0.8510 baseline, the claimed +3.25% mean gain over the baseline shrinks or disappears.

Watch

Extended reading notes

Core claim

On the paper's own terms, the core claim is that the two inductive biases in heterogeneous graph transformers — parameter-efficient cross-type transfer and relation-aware specialization — are complementary, not competing, and can be fused through a mechanism that gives each branch its own type-conditioned feature-wise gate before summation (TC-NGSF). A second claim is that relative time can be embedded directly into the attention kernel by rotating query and key vectors by opposite half-phases of the edge-wise time interval (RoTA), and that this temporal mechanism contributes more to the aggregate gain than the additive temporal encoding used by the baseline. The evidence is an ablation matr

Load-bearing premise

All reported improvements are relative to the authors' own reimplementation of the baseline on their preprocessing pipeline, and if that reimplementation is not as strong as a properly reproduced original, every relative gain is inflated.

Editorial extensions

If this is right

  • If the claims hold, temporal heterogeneous graph models do not have to choose between parameter-efficient sharing and relation-aware specialization; TC-NGSF provides a concrete, parameter-light way to fuse both.
  • RoTA demonstrates that relative time can be embedded in attention scores via rotation, opening the door to time-aware scoring beyond additive sinusoidal features.
  • The RTTA branch's lower per-layer edge cost means the accuracy is obtainable at a smaller computational footprint, which is relevant at billion-edge scales.
  • The consistency of the +3.25% mean gain across six tasks and four datasets says the recipe generalizes, not just fits one benchmark.

Reading between the lines

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

  • The paper compares against HGT under its own preprocessing pipeline; Table 2 shows that pipeline alone moves HGT's PF-L1 NDCG by roughly 13 points, an order of magnitude larger than the model-level gain. We infer that the +3.25% figure should be read as an improvement over the authors' own HGT setup; an externally reproduced baseline could shift the headline number.
  • RoTA is a temporal analogue of rotary position embeddings, so the half-phase rotation idea could transfer to any setting with relative displacement — continuous-time event sequences, trajectory data, or non-graph sequence models — though the paper does not test these.
  • The paper flags an induced-time inconsistency for nodes reached via non-timestamped edges under minibatch sampling; a globally consistent time assignment is a natural follow-up that could further tighten the temporal modeling.
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 THGFM, a dual-branch temporal heterogeneous graph transformer. One branch (SSTA) uses type-specific input adapters followed by shared Q/K/V projections for cross-type transfer; the other (RTTA) retains HGT-style type- and relation-specific parameters. The branches are fused via DRSF, instantiated with TC-NGSF, a type-conditioned, feature-wise gating mechanism that avoids competitive normalization. Relative time is injected into attention by RoTA, a rotary scheme that rotates queries and keys by half-phases of edge time. The model is evaluated on OAG-CS, OGBN-MAG, HTAG-ArXiv, and HTAG-DBLP against HGT, CTRL, ieHGCN, and SeHGNN under a fixed, disclosed hyperparameter grid, with ablations over fusion operators and temporal encodings. The abstract claims a +3.25% six-task mean gain and per-task peak gains up to +12.37%.

Significance. If the empirical claims were fully supported, THGFM would be a useful contribution: it directly addresses a real design tension between shared and specialized heterogeneous attention, and it makes relative time part of the attention score rather than only an additive feature. The experimental protocol is careful in several respects: hyperparameters are fixed to a disclosed grid to avoid tuning confounds; all models share preprocessing, sampling, and training procedures; nine runs over three seeds are reported with deterministic CUDA; and Table 2 explicitly quantifies the effect of preprocessing instead of hiding it. The architecture equations are explicit and there are no circular derivations or fitted constants. However, the headline results as written are not a faithful summary of the tables: the +3.25% mean and the listed peak gains come from different configurations, and the HGT baseline, while strong after the authors' preprocessing, is not externally anchored. These are fixable reporting and validation issues, but they are load-bearing for the central claim.

major comments (4)
  1. [Abstract; §5.4, Tables 3–4] The abstract states that THGFM delivers a +3.25% six-task mean gain, with peak gains of +12.37%, +4.87%, +1.18%, +4.24%, +3.73%, and +4.61%. In Table 4(b), the +3.25% figure is the TC-NGSF row averaged over the three temporal configurations, whereas the per-task peaks are taken from the 'Best single (operator, configuration)' row and are achieved by different operators/configs: SF@r110_s110 for PV, NGLF@r111_s111 for PF-L2, NGSF@r110_s110 for OGBN-MAG, etc. Under TC-NGSF, the PF-L2 gain is only +1.23% and PV is +9.04%. Thus the abstract presents an aggregate from one configuration together with maxima from other configurations. Moreover, the word 'consistently outperforms' is not supported by Table 3, where many rows are negative relative to HGT (e.g., RTTA r001_s001 PF-L2 -5.17; NGLF r111_s111 OGBN-MAG -4.56). Please report a single configuration's full six-task results or clearly label
  2. [§5.2, Table 2] All relative gains are computed against a single HGT row produced by the authors' reimplementation on their own preprocessing pipeline. Table 2 shows that changing only the preprocessing pipeline moves HGT's PF-L1 NDCG from 0.718±0.014 to 0.8510±0.013 — an absolute shift of about 0.133 NDCG points, which is larger in relative terms than the entire +3.25% model-level gain. The paper gives no external anchor: no canonical HGT checkpoint, no reproduction of published HGT scores on these tasks/splits, and no sensitivity analysis of HGT to the fixed hyperparameters under the new pipeline. Consequently, the reader cannot tell whether the '+X% vs HGT' cells are measured against a strong or an artificially weak baseline. Please add at least one external anchor or a hyperparameter sensitivity study for HGT on the authors' pipeline.
  3. [§5.2, Dataset Preprocessing] The text says the graphs are built with 'split-aware supervision' but does not define the term. If the preprocessing uses knowledge of which nodes are in train/validation/test to construct edges or features, this could leak supervision; if it does not, the wording should say so explicitly. Given the large preprocessing effect shown in Table 2, this ambiguity is material. Please specify exactly what information from the splits enters graph construction, and confirm that no validation/test labels or future timestamps are used during training.
  4. [§5.4, Table 4(a)] The text concludes that 'TC-NGSF under RoTA-only emerges as the most suitable THGFM configuration.' This is contradicted by Table 4(a): under r110_s110, NGSF has a six-task mean gain of +3.75% versus TC-NGSF's +3.48%, while using fewer parameters (9.38M vs 9.43M). If the selection criterion is the knee of the accuracy–cost curve, NGSF dominates TC-NGSF in that configuration. Please either correct the recommendation or justify a different selection criterion.
minor comments (5)
  1. [Table 4(a)] The 'Overall Mean Gain vs. HGT (%)' column appears to be the mean of the top-three entries only, not the mean over all operators. Please label it as such.
  2. [§3.1, Eq. (2)] The offset c is never concretely specified. Since RoTA's rotation angle is proportional to Δt, the chosen value of c can affect attention scores. State how c is set and, ideally, report sensitivity to c.
  3. [§5.2] The notation r001_s001 / r110_s110 / r111_s111 is not self-explanatory. Please define the bit positions explicitly (e.g., which bit controls RTE and which controls RoTA).
  4. [Abstract] Minor typographical issues: 'a adaptive' should be 'an adaptive'; there are also irregular spaces around 'THGFM' in the abstract and elsewhere.
  5. [§5.3] The paper reports standard deviations over nine runs but no paired tests or confidence intervals for the relative gains. A paired permutation test or bootstrap interval would strengthen the 'consistent outperformance' claim, especially on cells where the mean difference is smaller than the reported standard deviation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: THGFM's equations are explicit constructions and the +3.25% claim is an empirical benchmark summary, not a fitted constant or self-citation chain.

full rationale

The paper's derivation chain is architectural: SSTA, RTTA, RoTA, and TC-NGSF are defined by explicit equations (Eqs. 5–14), with no parameter fitted to the target metric and then renamed as a prediction. The +3.25% six-task mean gain is an equal-weight average of measured per-task relative gains vs. HGT (Tables 3–4), a summary of benchmark results on external datasets (OAG-CS, OGBN-MAG, HTAG), not a quantity entailed by the model definitions. No self-citations appear; RoFormer and HGT are cited as external prior work. Table 2's demonstration that preprocessing moves HGT's PF-L1 NDCG from 0.718 to 0.851 is a legitimate baseline-fairness concern, but it is not circularity: the paper discloses the protocol, reports standard deviations, and does not hide the fact that its HGT baseline uses the same preprocessing pipeline. The admitted induced-time inconsistency (Sec. 3.1, Sec. 6) is a stated limitation, not a circular reduction. There is no equation where an output equals an input by construction, and the central empirical claim is externally falsifiable. Therefore no circularity is present; any concerns about baseline anchoring or test-set model selection are correctness/robustness issues, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The learned network weights (input adapters, Q/K/V projections, A_r/M_r, gate MLPs, type embeddings) are model parameters, not ledger free parameters. The hand-set quantities conditioning the results are the temporal offset c (Eq. 2), the unspecified RoTA angle scaling (Eq. 4), and the fixed HGT-compatible hyperparameter grid. No invented entities (particles, forces, dimensions) are introduced; SSTA, RTTA, RoTA, and TC-NGSF are architectural modules. Axioms: the RoFormer rotation identity (standard math); the semantic meaningfulness of publication-year intervals (domain assumption); the per-batch induced-time protocol of Eq. (1), whose cross-batch inconsistency the paper concedes; and the transferability of HGT-tuned hyperparameters to all baselines.

free parameters (3)
  • Temporal offset c (Eq. 2) = unspecified non-negative constant
    Chosen by hand so Delta-t = t_i - t_j + c is non-negative; its value is never stated, so the absolute scale of relative times — and hence RoTA's rotation angles — is unreported.
  • RoTA rotation-angle scaling (Eq. 4) = unspecified
    The paper says theta_ji is proportional to Delta-t_ji but never gives the constant or frequency schedule; this mapping determines how strongly time reshapes attention and is needed to reproduce the model.
  • Fixed HGT-compatible hyperparameter grid = d=256, L=3, 8 heads, lr=1e-3, dropout=0.2, depth=6, width=128, 200 epochs
    Disclosed and fixed across all models to avoid tuning confounds; still hand-chosen values on which all reported results are conditioned.
assumptions (4)
  • standard math Rotary position embedding identity: rotating q by +theta/2 and k by -theta/2 shifts the dot product phase by theta
    RoTA's design rests on the RoFormer rotation property; this is an uncontroversial trigonometric identity, but it also means RoTA implements a specific band-limited kernel on relative time.
  • domain assumption Edge timestamps (publication years) are semantically meaningful relative times for attention
    RoTA's premise — papers with closer publication years deserve higher attention — is asserted in the Introduction ('temporal separation should reshape which neighbors receive high attention') but not verified against alternatives such as absolute time features.
  • domain assumption The induced node-time assignment in Eq. (1) is temporally consistent per batch
    Non-timestamped edges inherit the target context time; the paper acknowledges the same node receives different induced times across sampled batches (Section 3.1), so consistency holds only path-wise within a batch, and the Conclusion lists this as an open problem.
  • domain assumption HGT-compatible hyperparameters transfer fairly to all baselines (ieHGCN, SeHGNN, CTRL)
    A single fixed configuration is used for every model; if some baselines need different depths or widths to be competitive, the 'outperforms' comparison is skewed. Disclosed, but load-bearing for the central claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of THGFM: Dual-Branch Temporal Heterogeneous Graph Fusion Model." pith.science (2026). https://pith.science/paper/YI5HGQEB

@misc{pith2026260727303,
  author       = {Pith},
  title        = {Pith review of: THGFM: Dual-Branch Temporal Heterogeneous Graph Fusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YI5HGQEB}},
  note         = {Machine review of arXiv:2607.27303}
}
abstract

Temporal heterogeneous graphs offer a natural abstraction for dynamic relational systems in which diverse node and relation types co-exist and evolve over time. Learning on such graphs requires jointly modeling cross-type structural heterogeneity and the temporal dynamics of interactions, yet existing methods still struggle to reconcile parameter-efficient cross-type transfer with relation-aware specialization, and typically inject time only as additive features outside the attention kernel. We propose \textbf{THGFM}, a web-scale temporal heterogeneous graph fusion model that addresses both limitations within a unified dual-path architecture. THGFM couples a \textit{Shared-Space Temporal Attention} branch for parameter-efficient cross-type transfer with a \textit{Relational Type-Partitioned Temporal Attention} branch for relation-aware specialization, and integrates them through \textit{Dual-Path Relational--Shared Fusion}, instantiated with \textit{Type-Conditioned Non-Competitive Gated Sum Fusion}: a adaptive mechanism that assigns independent, type-conditioned feature-wise gates to the shared and specialized branches, allowing both to be amplified or suppressed without zero-sum competition. To directly incorporate relative time into the attention score, THGFM further introduces \textit{Rotary Temporal Attention}, which rotates queries and keys by half-phases of relative time before matching. THGFM consistently outperforms baseline graph transformer models on academic graphs benchmarks, delivering a $+3.25\%$ six-task mean gain, with peak relative gains of $+12.37\%$ on OAG-CS PV, $+4.87\%$ on PF-$L_2$, and $+1.18\%$ on PF-$L_1$, and $+4.24\%$, $+3.73\%$, and $+4.61\%$ on OGBN-MAG, HTAG-ArXiv, and HTAG-DBLP, respectively.

Figures

Figures reproduced from arXiv: 2607.27303 by the authors.

Figure 1
Figure 1. OAG as a temporal heterogeneous graph with multiple node types and re￾lation semantics that evolve over time. Timestamp availability is relation-dependent: publication-related edges are time-stamped by publication year, while author-affiliation links in this example do not provide explicit timestamps and are therefore treated as edges with missing time (ψ(e) = ⊥). The figure is adapted from [9]. Representative examp… view at source ↗
Figure 2
Figure 2. Overall architecture of THGFM. Each layer executes two temporal heteroge￾neous attention branches—SSTA (Shared-Space Temporal Attention) and RTTA (Re￾lational Type- Partitioned Temporal Attention)—both modulated by Rotary Temporal Attention (RoTA), and fuses their outputs through Dual-Path Relational–Shared Fu￾sion (DRSF) before passing to the next layer via a gated residual connection. Representations are then upda… view at source ↗
Figure 3
Figure 3. Two THGFM attention branches used by DRSF. SSTA provides parameter￾efficient cross-type transfer in a shared space, while RTTA preserves type- and relation￾aware specialization; both apply RoTA before attention scoring. 4.2 Dual-Path Relational–Shared Fusion (DRSF) DRSF combines two branches with complementary inductive roles: SSTA for shared-space cross-type transfer and RTTA for relation-aware heterogeneous in￾ter… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: DRSF. Branch outputs z ssta i and z rtta i are activated with GELU and fed into one of six fusion operators. Gated fusion operators (NGLF, NGSF, TC-NGLF, TC￾NGSF) compute non-competitive sigmoid gates from concatenated branch features— optionally conditioned on a targe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 3 canonical work pages

  1. [1]

    arXiv preprint arXiv:2602.04768 (2026), https://arxiv.org/abs/2602.04768

    Bechler-Speicher, M., Gottlieb, Y., Isakov, A., Abensur, D., Tavory, A., Haimovich, D., Guy, I., Weinsberg, U.: Billion-scale graph foundation models. arXiv preprint arXiv:2602.04768 (2026), https://arxiv.org/abs/2602.04768

  2. [2]

    Natural Language Engineering23(1), 155–162 (2017), https://doi.org/10.1017/S1351324916000334

    Church, K.W.: Word2vec. Natural Language Engineering23(1), 155–162 (2017), https://doi.org/10.1017/S1351324916000334

  3. [3]

    arXiv preprint arXiv:2110.13889 (2021), https://arxiv.org/abs/2110

    Fan, Y., Ju, M., Zhang, C., Zhao, L., Ye, Y.: Heterogeneous temporal graph neu- ral network. arXiv preprint arXiv:2110.13889 (2021), https://arxiv.org/abs/2110. 13889

  4. [4]

    arXiv preprint arXiv:2405.00476 (2024), https://arxiv.org/ abs/2405.00476

    Feng, Z., Wang, R., Wang, T., Song, M., Wu, S., He, S.: A comprehensive sur- vey of dynamic graph neural networks: Models, frameworks, benchmarks, experi- ments and challenges. arXiv preprint arXiv:2405.00476 (2024), https://arxiv.org/ abs/2405.00476

  5. [5]

    In: Proceedings of The Web Confer- ence 2020

    Fu, X., Zhang, J., Meng, Z., King, I.: MAGNN: Metapath aggregated graph neural network for heterogeneous graph embedding. In: Proceedings of The Web Confer- ence 2020. pp. 2331–2341 (2020), https://doi.org/10.1145/3366423.3380297

  6. [6]

    ACM Computing Surveys54(4), 1–37 (2021), https://doi.org/ 10.1145/3447772

    Hogan, A., Blomqvist, E., Cochez, M., d’Amato, C., de Melo, G., Gutierrez, C., Gayo, J.E.L., Kirrane, S., Neumaier, S., Polleres, A., Navigli, R., Ngomo, A.C.N., Rashid, S.M., Rula, A., Schmelzeisen, L., Sequeda, J., Staab, S., Zimmermann, A.: Knowledge graphs. ACM Computing Surveys54(4), 1–37 (2021), https://doi.org/ 10.1145/3447772

  7. [7]

    In: Proceedings of the AAAI 18 Y

    Hong, H., Guo, H., Lin, Y., Yang, X., Li, Z., Ye, J.: An attention-based graph neural network for heterogeneous structural learning. In: Proceedings of the AAAI 18 Y. Peng et al. Conference on Artificial Intelligence. vol. 34, pp. 4132–4139 (2020), https://doi. org/10.1609/aaai.v34i04.5833

  8. [8]

    arXiv preprint arXiv:2005.00687 (2020), https://arxiv.org/abs/2005.00687

    Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687 (2020), https://arxiv.org/abs/2005.00687

Show all 30 references
  1. [9]

    In: Pro- ceedings of The Web Conference 2020

    Hu, Z., Dong, Y., Wang, K., Sun, Y.: Heterogeneous graph transformer. In: Pro- ceedings of The Web Conference 2020. pp. 2704–2710 (2020), https://doi.org/10. 1145/3366423.3380027

  2. [10]

    Knowledge-Based Systems339, 115514 (2026), https://www

    Li, C., Xie, Y., Yu, C., Zhao, J., Xia, Y., Kong, B., Li, Z., Niu, D.: CTRL: Continuous-time representation learning on temporal heterogeneous informa- tion network. Knowledge-Based Systems339, 115514 (2026), https://www. sciencedirect.com/science/article/pii/S095070512600256X

  3. [11]

    Data Science and Engineering8, 98–111 (2023), https://doi.org/10

    Li, L., Duan, L., Wang, J., He, C., Chen, Z., Xie, G., Deng, S., Luo, Z.: Memory-enhanced transformer for representation learning on temporal heteroge- neous graphs. Data Science and Engineering8, 98–111 (2023), https://doi.org/10. 1007/s41019-023-00207-w

  4. [12]

    In: 2024 IEEE 40th International Conference on Data Engineering (ICDE)

    Liang, Y., Zhang, W., Sheng, Z., Yang, L., Jiang, J., Tong, Y., Cui, B.: HGAMLP: Heterogeneous graph attention MLP with de-redundancy mechanism. In: 2024 IEEE 40th International Conference on Data Engineering (ICDE). pp. 2779–2791 (2024), https://doi.org/10.1109/ICDE60146.2024.00217

  5. [13]

    Journal of King Saud University Computer and Information Sciences37, 191 (2025), https://doi.org/ 10.1007/s44443-025-00222-z

    Liu, K., Li, X., Liu, L., Xv, Z., Pei, A., Ji, R.: AFMF: Adaptive fusion of multi- hop neighborhood features in graph convolutional network. Journal of King Saud University Computer and Information Sciences37, 191 (2025), https://doi.org/ 10.1007/s44443-025-00222-z

  6. [14]

    arXiv preprint arXiv:2412.08937 (2024), https://arxiv.org/abs/2412.08937

    Liu, Y., Xie, Q., Shi, J., Shen, J., He, T.: Multi-scale heterogeneous text-attributed graph datasets from diverse domains. arXiv preprint arXiv:2412.08937 (2024), https://arxiv.org/abs/2412.08937

  7. [15]

    In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Lv, Q., Ding, M., Liu, Q., Chen, Y., Feng, W., He, S., Zhou, C., Jiang, J., Dong, Y., Tang, J.: Are we really making much progress? revisiting, benchmarking, and refining heterogeneous graph neural networks. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discov...

  8. [16]

    In: Proceedings of the 26th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining

    Niu, X., Li, B., Li, C., Xiao, R., Sun, H., Deng, H., Chen, Z.: A dual heteroge- neous graph attention network to improve long-tail performance for shop search in e-commerce. In: Proceedings of the 26th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining. pp. 2025–2...

  9. [17]

    arXiv preprint arXiv:2006.10637 (2020), https://arxiv.org/abs/2006.10637

    Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., Bronstein, M.M.: Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637 (2020), https://arxiv.org/abs/2006.10637

  10. [18]

    Schlichtkrull, M., Kipf, T.N., Bloem, P., van den Berg, R., Titov, I., Welling, M.: Modelingrelational datawith graphconvolutional networks.In: TheSemanticWeb – 15th International Conference, ESWC 2018. pp. 593–607 (2018). https://doi.org/ 10.1007/978-3-319-93417-4_38, https:/...

  11. [19]

    In: Proceedings of the 2023 International Conference on Machine Learning and Cybernetics

    Shang, Y., Ren, Y., Peng, H., Wang, Y., Wang, G., Li, Z.C., Yang, Y., Li, Y.: A perspective survey on industrial knowledge graphs: Recent advances, open chal- lenges, and future directions. In: Proceedings of the 2023 International Conference on Machine Learning and Cybernetic...

  12. [20]

    arXiv preprint arXiv:2104.09864 (2021), https://arxiv.org/abs/2104.09864

    Su, J., Lu, Y., Pan, S., Murtadha, A., Wen, B., Liu, Y.: RoFormer: Enhanced trans- former with rotary position embedding. arXiv preprint arXiv:2104.09864 (2021), https://arxiv.org/abs/2104.09864

  13. [21]

    In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=BylA_C4tPr

    Vashishth, S., Sanyal, S., Nitin, V., Talukdar, P.: Composition-based multi- relational graph convolutional networks. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=BylA_C4tPr

  14. [22]

    Quantitative Science Studies1(1), 396–413 (2020), https://doi.org/10.1162/qss_a_00021

    Wang, K., Shen, Z., Huang, C., Wu, C.H., Dong, Y., Kanakia, A.: Microsoft Aca- demic Graph: When experts are not enough. Quantitative Science Studies1(1), 396–413 (2020), https://doi.org/10.1162/qss_a_00021

  15. [23]

    In: Proceedings of The World Wide Web Conference

    Wang, X., Ji, H., Shi, C., Wang, B., Ye, Y., Cui, P., Yu, P.S.: Heterogeneous graph attention network. In: Proceedings of The World Wide Web Conference. pp. 2022–2032 (2019), https://doi.org/10.1145/3308558.3313562

  16. [24]

    In: International Conference on Learning Represen- tations (2020), https://openreview.net/forum?id=rJeW1yHYwH

    Xu, D., Ruan, C., Korpeoglu, E., Kumar, S., Achan, K.: Inductive representation learning on temporal graphs. In: International Conference on Learning Represen- tations (2020), https://openreview.net/forum?id=rJeW1yHYwH

  17. [25]

    In: Proceedings of the AAAI Conference on Artificial Intel- ligence

    Yang, X., Yan, M., Pan, S., Ye, X., Fan, D.: Simple and efficient heterogeneous graph neural network. In: Proceedings of the AAAI Conference on Artificial Intel- ligence. vol. 37, pp. 10816–10824 (2023), https://doi.org/10.1609/aaai.v37i9.26283

  18. [26]

    IEEE Transactions on Knowledge and Data Engineering35(2), 1637–1650 (2023), https://doi.org/10.1109/TKDE.2021

    Yang, Y., Guan, Z., Li, J., Zhao, W., Cui, J., Wang, Q.: Interpretable and efficient heterogeneous graph convolutional network. IEEE Transactions on Knowledge and Data Engineering35(2), 1637–1650 (2023), https://doi.org/10.1109/TKDE.2021. 3101356

  19. [27]

    In: Advances in Neural Information Processing Systems

    Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., Le, Q.V.: XLNet: Generalized autoregressive pretraining for language understanding. In: Advances in Neural Information Processing Systems. vol. 32 (2019), https://papers.nips.cc/ paper/2019/hash/dc6a7e655d7e5840e66...

  20. [28]

    IEEE Transactions on Knowledge and Data Engineering35(6), 5935–5947 (2023), https://doi.org/10.1109/TKDE.2022

    Yu, L., Sun, L., Du, B., Liu, C., Lv, W., Xiong, H.: Heterogeneous graph repre- sentation learning with relation awareness. IEEE Transactions on Knowledge and Data Engineering35(6), 5935–5947 (2023), https://doi.org/10.1109/TKDE.2022. 3160208

  21. [29]

    In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    Zhang, F., Liu, X., Tang, J., Dong, Y., Yao, P., Zhang, J., Gu, X., Wang, Y., Shao, B., Li, R., Wang, K.: OAG: Toward linking large-scale heterogeneous entity graphs. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 25...

  22. [30]

    arXiv preprint arXiv:2408.09697 (2024)

    Zhong, Y., Su, J., Wu, C., Wang, M.: Heta: Distributed training of heterogeneous graph neural networks. arXiv preprint arXiv:2408.09697 (2024)

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.