Pith. sign in

REVIEW 4 major objections 7 minor 24 references

IMPA-HGAE:Intra-Meta-Path Augmented Heterogeneous Graph Autoencoder

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A self-supervised heterogeneous graph autoencoder that mines the intermediate nodes of meta-paths outperforms baselines on four academic graphs.

desk verdict Plausible extension of masked graph autoencoders, but the ablation supporting the main mechanism is unreadable and the empirical claims are overstated. read the letter →

arxiv 2506.06809 v1 pith:VKKNXPCI submitted 2025-06-07 cs.LG cs.AI

Di Lin , Wanjing Ren , Xuanbin Li , Rui Zhang This is my paper
classification cs.LGcs.AI
keywords heterogeneousgraphneuralnetworkself-supervisedlearningautoencodermeta-pathmaskedreconstructionnodeclassificationattentionmechanismrepresentation
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

This paper tries to establish that a heterogeneous graph autoencoder can learn better node embeddings by using the nodes inside a meta-path, not just the two endpoints. The proposed model, IMPA-HGAE, masks and reconstructs both node features and meta-path adjacency matrices, and enhances target-node embeddings by attention-based propagation of intermediate-node information along each meta-path. If this is right, generative self-supervised learning on heterogeneous graphs no longer needs to discard the semantics of non-target nodes, and masked autoencoders can capture longer-range structure without labeled data. The paper reports node-classification gains over ten baselines on DBLP, Freebase, ACM, and AMiner.

What carries the argument

The carrying mechanism is subgraph-wise, meta-path-guided attention propagation: after masking, subgraphs are extracted, non-target node embeddings are aligned with target embeddings by an MLP ($H_o = \sigma^{(2)}(W^{(2)}\sigma^{(1)}(W^{(1)}X_o+b^{(1)})+b^{(2)})$), and stacked generalized GAT layers compute node-level attention coefficients $\alpha^{\Phi}_{ij}$ to aggregate messages from intermediate nodes into terminal nodes. A second load-bearing component is the dual masked reconstruction objective, combining a feature loss and a meta-path adjacency loss, which drives the encoder to preserve both attributes and long-range structure. The masking policy itself matters: degree-based edge masking, defined through averaged in- and out-degree weights, outperforms random and attention-score-based masking in the reported experiments.

What would settle it

Run IMPA-HGAE with the intra-meta-path propagation module removed but all masking and reconstruction objectives kept. If node-classification accuracy on DBLP and ACM stays within noise of the full model, the claim that intermediate-node information drives the gains would be refuted.

Watch

Extended reading notes

Core claim

The central discovery the paper claims is that the information lost by existing meta-path-based models is concentrated in the intermediate, non-target nodes, and that recovering it improves learned representations. IMPA-HGAE encodes masked features and meta-paths with a graph encoder, projects non-target node features into the target embedding space with an MLP, then runs stacked, type-aware attention-based message passing from intermediate to terminal nodes inside each meta-path subgraph. The decoder reconstructs both the masked feature matrix and the masked meta-path adjacency, with cosine-similarity losses scaled by exponents, and both objectives are optimized jointly. In node-classification experiments, the model outperforms the compared baselines on most metric/dataset/split combinations, and ablation results indicate that the intra-meta-path propagation and the joint reconstruction each contribute to the gain.

Load-bearing premise

Intermediate nodes along a meta-path carry signal useful for target-node classification, and a single MLP projection aligns them well enough that attention-based message passing adds information instead of noise.

Editorial extensions

If this is right

  • Masked autoencoders for heterogeneous graphs should model intermediate nodes on meta-paths, not only endpoints, to capture local proximity and semantic information.
  • Degree-based meta-path masking produces larger gains than random or attention-score-based masking, suggesting that targeted corruption of dense edge regions is a useful design choice.
  • Joint reconstruction of features and meta-path adjacency matrices outperforms either reconstruction objective alone.
  • The learned embeddings transfer to node classification, improving Micro-F1, Macro-F1, and AUC over the compared baselines on DBLP, Freebase, ACM, and AMiner in most settings.
  • The paper's decoder visualization shows imperfect reconstruction, which points to stronger decoders or auxiliary tasks as a natural next step.

Reading between the lines

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

  • The same intra-meta-path mechanism should be tested on non-academic heterogeneous graphs, such as e-commerce or knowledge graphs, where intermediate nodes carry rich attributes; the paper does not report such runs.
  • The approximate inverse-softmax recovery used for attention-score-based masking could be replaced by storing pre-softmax logits during encoding, a fix the paper itself flags as unresolved.
  • The degree-based masking result implies a testable rule: corrupt high-degree edge regions to regularize heterogeneous graphs, which may transfer to other masked graph autoencoders.
  • Because the decoder reconstruction is unsatisfactory, grafting a stronger decoder or auxiliary reconstruction targets onto IMPA-HGAE may yield further classification gains beyond those reported.
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 / 7 minor

Summary. The paper proposes IMPA-HGAE, a self-supervised heterogeneous graph autoencoder. It masks and reconstructs both node feature matrices and meta-path adjacency matrices, introduces three meta-path masking strategies (random, degree-based, and attention-score-based via an inverse softmax transformation), and augments target-node embeddings by propagating information from intermediate non-target nodes along meta-paths. The model is evaluated on node classification over DBLP, Freebase, ACM, and AMiner and compared against seven baselines. The abstract claims that IMPA-HGAE 'achieves superior performance on heterogeneous datasets' and that the gains come from exploiting internal node information along meta-paths.

Significance. If the central claim were fully supported, the paper would make a useful contribution: the idea of using intermediate nodes along meta-paths in a generative self-supervised framework is a reasonable extension of masked autoencoding to heterogeneous graphs, and the systematic comparison of masking policies is relevant to the graph SSL community. The manuscript is also commendable for evaluating on four datasets with three training splits and for candidly discussing the limitation of the inverse-softmax approximation in Section 4.3. However, the paper as submitted does not consistently support its headline claim of superiority: Table 1 contains multiple cells where baselines outperform IMPA-HGAE, and the ablation evidence for the novel mechanism is unreadable because Figure 4 consists of placeholder tokens rather than a plot. No code or reproducibility artifacts are provided. The significance is therefore conditional on a corrected empirical presentation and a more cautious interpretation of the results.

major comments (4)
  1. [Abstract and Table 1] The abstract's claim that IMPA-HGAE 'achieves superior performance on heterogeneous datasets' is not supported by Table 1. For example, on DBLP Mi-F1@20 IMPA-HGAE scores 92.09 while DiffGraph scores 92.90 and HGMAE scores 92.71; on DBLP Ma-F1@20 IMPA-HGAE scores 91.64 while HGMAE scores 92.28; on Freebase Ma-F1@20 IMPA-HGAE scores 62.04 while DiffGraph scores 62.93 and Ma-F1@40 scores 64.54 while DiffGraph scores 64.87; on ACM Ma-F1@60 IMPA-HGAE scores 91.39 while DiffGraph scores 91.73; and on AMiner AUC@20 IMPA-HGAE scores 92.40 while HGMAE scores 93.22. In addition, many of the reported wins are within one standard deviation of the baseline results. The empirical claim should be qualified to state that IMPA-HGAE is competitive or achieves best results on a subset of metrics and splits, with statistical significance assessed rather than asserted.
  2. [Section 4.2, Figure 4] The ablation study that is supposed to isolate the contribution of the intra-meta-path propagation mechanism is not verifiable in the submitted manuscript. The text in Section 4.2 states that degree-based masking gives the largest improvement and that 'incorporating neighboring node enhancement significantly improves model performance,' but Figure 4, the figure cited for these claims, contains a long sequence of placeholder tokens (e.g., '/uni00000031/uni00000052/...') rather than a readable plot. No numeric ablation table is provided. Without a readable figure or a table of ablation numbers, the central claim that the intra-meta-path component adds signal beyond the masked-reconstruction objective shared with HGMAE is unsupported.
  3. [Section 3.2, Eq. (4)] The load-bearing assumption of the proposed augmentation is that a single MLP can map the features of all non-target node types into the same space as target-node embeddings, so that attention-based message passing from intermediate nodes to terminal nodes adds useful information rather than noise. Eq. (4) defines this mapping, but the paper provides no direct evidence that the alignment is effective, and the ablation as presented cannot separate the benefit of this propagation from the benefit of the dual masking and reconstruction objective. Since HGMAE already uses masked reconstruction of meta-paths and node features, the paper needs a controlled comparison (e.g., IMPA-HGAE without the propagation module, or with randomly projected intermediate features) to substantiate the claim that the improvement comes specifically from 'fully exploiting internal node information along meta-paths.'
  4. [Section 4.3, Eq. (3)] The authors correctly note in Section 4.3 that the inverse-softmax recovery in Eq. (3) preserves only relative magnitudes because of the translation invariance of softmax, meaning the recovered attention scores are not uniquely determined. This is not fatal to the main empirical result because degree-based masking, not attention-score-based masking, yields the best performance, but it does undermine the interpretability analysis that uses attention-score masking to compare masking policies. The paper should either provide a concrete justification for choosing a particular constant C or present the attention-score-based results as an approximation with a sensitivity analysis over C.
minor comments (7)
  1. [Section 4.2] The text says 'see the Figure 3' when describing the ablation results, but the ablation is presented as Figure 4; the figure citation should be corrected.
  2. [Figure 4 caption] The caption 'AblationStudyonMaskingStrategiesandModelComponentsonACMDataset' is missing spaces and is difficult to parse; the figure itself needs to be replaced with a legible plot or replaced by a table.
  3. [Eq. (9)] The function δ in Eq. (9) is not defined. If it denotes an activation such as sigmoid, it should be stated explicitly.
  4. [Section 4.1] The paper says comparisons are made with ten baseline methods, but Table 1 lists only seven baselines: HetGNN, HAN, DMGI, HeCo, GraphMAE, HGMAE, and DiffGraph. The list of baselines and the table should be reconciled.
  5. [Reference [24]] The DiffGraph reference contains placeholder text '[Conference Name]' and '[Volume Number]'; this needs to be completed.
  6. [Section 2] The definition of a heterogeneous graph in Section 2 contains a repeated phrase 'represents the set of nodes' and appears to have a typo in the listing of components; the notation should be cleaned up.
  7. [Section 5] The sentence discussing Figure 6(c) says the reconstruction is 'not entirely satisfactory,' which is in tension with the paper's overall positive framing; the authors should clarify whether this is a limitation or a direction for future work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims rest on external benchmark comparisons and architectural loss definitions, not on fitted values renamed as predictions or on a self-citation chain.

full rationale

The paper's derivation chain consists of architectural choices (masked reconstruction of feature matrices and meta-path adjacency matrices, Eqs. 8-11) and an attention-based propagation scheme for intermediate meta-path nodes (Eqs. 4-7). These are explicit model definitions with reconstruction objectives, and the headline result is an empirical node-classification comparison against published baselines on DBLP, Freebase, ACM, and AMiner. The downstream metrics are computed on held-out splits, so the reported numbers are not fitted to the evaluation target. The inverse-softmax recovery in Eq. 3 is explicitly flagged by the authors as approximate due to softmax translation invariance, which is an acknowledged limitation rather than a circular step. The degree-based masking discussion in Section 4.3 is post-hoc reasoning about an observed ablation result, not a fitted parameter presented as a prediction; the claim that degree-based masking performs best is an experimental observation, and the subsequent uniformity hypothesis is speculative but does not reduce the result to its input. No load-bearing argument depends on a self-citation: the potentially same-name citation in Reference 15 is not used to justify the method, and no uniqueness theorem or external authority is invoked to forbid alternatives. Concerns that Table 1 does not uniformly show superiority and that Figure 4 appears corrupted in the supplied text are evidence-quality and correctness issues, not circularity. The paper is therefore self-contained with respect to its empirical evaluation, and no specific reduction of a claimed prediction to its own construction can be exhibited.

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

The central claim rests on a standard masked-reconstruction setup whose hyperparameters are not reported, plus three assumptions: that meta-path adjacency matrices are a sufficient structural prior, that intermediate meta-path nodes carry useful signal after a shallow MLP projection, and that inverse softmax with a constant approximately recovers attention scores. No invented entities are introduced; 'intra-meta-path propagation' is a mechanism, not a postulate.

free parameters (4)
  • mask_rate = not reported
    Fraction of meta-path edges or feature entries masked; Figure 5 sweeps 0.1-0.9 but the per-dataset chosen values and whether they are tuned on validation are not stated.
  • alpha_phi = not reported
    Per-meta-path attention weight in the L_mp loss (Eq. 11), inherited from HAN; the paper does not describe how it is computed or updated beyond 'HAN attention coefficients'.
  • gamma_1 and gamma_2 = not reported
    Scaling exponents in the cosine feature and meta-path losses (Eqs. 8 and 10); values are never stated.
  • inverse-softmax constant C = not reported
    Used in Eq. 3 to recover pre-softmax attention scores for attention-based masking; its value is not given even though the recovery quality depends on it.
assumptions (3)
  • domain assumption Meta-path adjacency matrices A_phi encode sufficient structural semantics for heterogeneous self-supervised learning.
    The pipeline masks and reconstructs A_phi (Eqs. 9-11) as a core objective; if the predefined meta-paths are uninformative, the objective is vacuous. Invoked throughout Section 3.1.
  • domain assumption Intermediate nodes along a meta-path carry signal useful for target-node classification, and a single MLP projection (Eq. 4) aligns their features with target embeddings.
    This is the core of the intra-meta-path propagation in Section 3.2; if false, the augmentation adds noise and the claimed gains would vanish.
  • ad hoc to paper The inverse-softmax transformation z_i = ln(p_i) + C recovers usable pre-normalized attention scores.
    Eq. 3. The paper itself states in Section 4.3 that softmax translation invariance means only relative magnitudes may be preserved, so the recovered scores are approximate by the authors' own account.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IMPA-HGAE:Intra-Meta-Path Augmented Heterogeneous Graph Autoencoder." pith.science (2026). https://pith.science/paper/VKKNXPCI

@misc{pith2026250606809,
  author       = {Pith},
  title        = {Pith review of: IMPA-HGAE:Intra-Meta-Path Augmented Heterogeneous Graph Autoencoder},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKKNXPCI}},
  note         = {Machine review of arXiv:2506.06809}
}
read the original abstract

Self-supervised learning (SSL) methods have been increasingly applied to diverse downstream tasks due to their superior generalization capabilities and low annotation costs. However, most existing heterogeneous graph SSL models convert heterogeneous graphs into homogeneous ones via meta-paths for training, which only leverage information from nodes at both ends of meta-paths while underutilizing the heterogeneous node information along the meta-paths. To address this limitation, this paper proposes a novel framework named IMPA-HGAE to enhance target node embeddings by fully exploiting internal node information along meta-paths. Experimental results validate that IMPA-HGAE achieves superior performance on heterogeneous datasets. Furthermore, this paper introduce innovative masking strategies to strengthen the representational capacity of generative SSL models on heterogeneous graph data. Additionally, this paper discuss the interpretability of the proposed method and potential future directions for generative self-supervised learning in heterogeneous graphs. This work provides insights into leveraging meta-path-guided structural semantics for robust representation learning in complex graph scenarios.

Figures

Figures reproduced from arXiv: 2506.06809 by the authors.

Figure 1
Figure 1. Comparative Analysis: Real-World Challenges in Heterogeneous Graphs (Left) vs. IMPA-HGAE’s Innovative Solutions (Right) (Left) Conventional approach: (C1) APCPA metapath illustration, (C2) Generic neigh￾borhood aggregation, (C3) Intermediate node-agnostic propagation. (Right) IMPA-HGAE framework: (C1) Annotated metapath, (C2) Author-to-paper re￾lation learning, (C3) Paper-to-affiliation attribute transfer. Self-supe… view at source ↗
Figure 2
Figure 2. The overall framework of IMPA-HGAE comprises four key modules: data pre￾processing, encoding, propagation, and decoding. Initially, the model performs random masking on the feature matrix while applying a degree-based masking strategy to the adjacency matrix. The preprocessed graph data is then encoded via a Graph Neu￾ral Network (GNN) to generate node embeddings. During the propagation phase, an attention mechanism… view at source ↗
Figure 3
Figure 3. Edge Attention Visualization Diagram As illustrated in the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Ablation Study on Masking Strategies and Model Components on ACM Dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Mask Analysis Diagram: (a)(b) Connected component trends under various masking ratios and methods; (c) 3D surface comparison between feature masking and meta-path masking. 4.3 Mask Policy Analysis According to the left figure in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Visualization of DBLP Dataset Distribution via t-SNE Dimensionality Reduc￾tion this study. Experimental results indicate that IMPA-HGAE significantly outper￾forms various baseline models. Overall, IMPA-HGAE introduces new advance￾ments to self-supervised generative lea…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [1]

    arXiv preprint arXiv:2202.07082 (2022)

    Zheng, X., Wang, Y., Liu, Y., et al.: Graph neural networks for graphs with heterophily: A survey. arXiv preprint arXiv:2202.07082 (2022). https://arxiv.org/abs/2202.07082

    arXiv 2022

  2. [2]

    In: Dong, Y., Wang, K., Sun, Y

    Hu, Z., Dong, Y., Wang, K., Sun, Y.: Heterogeneous graph transformer. In: Dong, Y., Wang, K., Sun, Y. (eds.) WWW 2020, pp. 2704–2710. ACM, New York (2020)

    work page 2020

  3. [3]

    In: Zhang, C., Metoyer, R., Chawla, N.V

    Tian, Y., Zhang, C., Metoyer, R., Chawla, N.V.: Recipe representation learning with networks. In: Zhang, C., Metoyer, R., Chawla, N.V. (eds.) CIKM 2021, pp. 1824–1833. ACM, New York (2021)

    work page 2021

  4. [4]

    In: Song, D., Huang, C., Swami, A

    Zhang, C., Song, D., Huang, C., Swami, A., Chawla, N.V.: Heterogeneous graph neural network. In: Song, D., Huang, C., Swami, A. (eds.) KDD 2019, pp. 793–803. ACM, New York (2019)

    work page 2019

  5. [5]

    In: IEEE ICDM 2022, pp

    Wang, B., Li, A., Pang, M., Li, H., Chen, Y.: GraphFL: A federated learning framework for semi-supervised node classification on graphs. In: IEEE ICDM 2022, pp. 498–507. IEEE, New York (2022)

    work page 2022

  6. [6]

    IEEE Transactions on Knowledge and Data Engineering 35(1), 857–876 (2021)

    Liu, X., Zhang, F., Hou, Z., Mian, L., Wang, Z., Zhang, J., Tang, J.: Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering 35(1), 857–876 (2021)

    work page 2021

  7. [7]

    arXiv preprint arXiv:1611.07308 (2016) 12 Di Lin, Wanjing Ren, Xuanbin Li, and Rui Zhang

    Kipf, T.N., Welling, M.: Variational graph auto-encoders. arXiv preprint arXiv:1611.07308 (2016) 12 Di Lin, Wanjing Ren, Xuanbin Li, and Rui Zhang

    arXiv 2016

  8. [8]

    In: KDD 2022, pp

    Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., Tang, J.: GraphMAE: Self-supervised masked graph autoencoders. In: KDD 2022, pp. 594–604. ACM, New York (2022)

    work page 2022

Show all 24 references
  1. [9]

    In: AAAI 2023, vol

    Tian, Y., Dong, K., Zhang, C., Zhang, C., Chawla, N.V.: Heterogeneous graph masked autoencoders. In: AAAI 2023, vol. 37, pp. 9997–10005. AAAI Press, Palo Alto (2023)

    2023

  2. [10]

    In: AAAI 2023, vol

    Cheng,J.,Li,M.,Li,J.,Tsung,F.:Wienergraphdeconvolutionalnetworkimproves graph self-supervised learning. In: AAAI 2023, vol. 37, pp. 7131–7139. AAAI Press, Palo Alto (2023)

    2023

  3. [11]

    In: The Web Conference 2023, pp

    Li, X., Ye, T., Shan, C., Li, D., Gao, M.: SeeGERA: Self-supervised semi-implicit graph variational auto-encoders with masking. In: The Web Conference 2023, pp. 143–153. ACM, New York (2023)

    2023

  4. [12]

    In: AAAI Conference on Artificial Intelligence, LNCS, vol

    Wang, L., Tao, X., Liu, Q., Wu, S.: Rethinking Graph Masked Autoencoders through Alignment and Uniformity. In: AAAI Conference on Artificial Intelligence, LNCS, vol. 38, pp. 15528–15536. Springer, Heidelberg (2024)

    2024

  5. [13]

    In: International Conference on Machine Learning, pp

    Abu-El-Haija, S., Perozzi, B., Kapoor, A., Alipourfard, N., Lerman, K., Harutyun- yan, H., Ver Steeg, G., Galstyan, A.: MixHop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In: International Conference on Machine Learning, pp. 21–29. PMLR (2019)

    2019

  6. [14]

    Advances in Neural Information Processing Systems33, 7793–7804 (2020)

    Zhu,J.,Yan,Y.,Zhao,L.,Heimann,M.,Akoglu,L.,Koutra,D.:Beyondhomophily in graph neural networks: Current limitations and effective designs. Advances in Neural Information Processing Systems33, 7793–7804 (2020)

    2020

  7. [15]

    In: European Conference on Computer Vision, pp

    Wang, J., Yan, S., Xiong, Y., Lin, D.: Motion guided 3D pose estimation from videos. In: European Conference on Computer Vision, pp. 764–780. Springer (2020)

    2020

  8. [16]

    In: Proceedings of the AAAI Conference on Artificial In- telligence, vol

    Bo, D., Wang, X., Shi, C., Shen, H.: Beyond low-frequency information in graph convolutional networks. In: Proceedings of the AAAI Conference on Artificial In- telligence, vol. 35, pp. 3950–3957 (2021)

    2021

  9. [17]

    Luan, S., Hua, C., Lu, Q., Zhu, J., Zhao, M., Zhang, S., Chang, X.W., Precup, D.: Is heterophily a real nightmare for graph neural networks to do node classification? arXiv preprint arXiv:2109.05641 (2021)

    2021 arXiv

  10. [18]

    arXiv preprint arXiv:2006.07988 (2020)

    Chien, E., Peng, J., Li, P., Milenkovic, O.: Adaptive universal generalized PageR- ank graph neural network. arXiv preprint arXiv:2006.07988 (2020)

    2020 arXiv

  11. [19]

    In: International Conference on Machine Learning, pp

    You, Y., Chen, T., Shen, Y., Wang, Z.: Graph contrastive learning automated. In: International Conference on Machine Learning, pp. 12121–12132. PMLR (2021)

    2021

  12. [20]

    In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp

    Wang, X., Liu, N., Han, H., Shi, C.: Self-supervised heterogeneous graph neural network with co-contrastive learning. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp. 1726–1736 (2021)

    2021

  13. [21]

    arXiv preprint arXiv:1911.08538 (2019)

    Ren, Y., Liu, B., Huang, C., Dai, P., Bo, L., Zhang, J.: Heterogeneous deep graph infomax. arXiv preprint arXiv:1911.08538 (2019)

    2019 arXiv

  14. [22]

    In: AAAI Conference on Artificial Intelligence, vol

    Park, C., Kim, D., Han, J., Yu, H.: Unsupervised attributed multiplex network embedding. In: AAAI Conference on Artificial Intelligence, vol. 34, no. 04, pp. 5371–5378. AAAI (2020)

    2020

  15. [23]

    In: The World Wide Web Conference, pp

    Wang, X., Ji, H., Shi, C., Wang, B., Ye, Y., Cui, P., Yu, P.S.: Heterogeneous graph attention network. In: The World Wide Web Conference, pp. 2022–2032. ACM (2019)

    2019

  16. [24]

    In: Proceedings of the [Conference Name], LNCS, vol

    Li, Z., Xia, L., Hua, H., Zhang, S., Wang, S., Huang, C.: DiffGraph: Heterogeneous Graph Diffusion Model (Extended Abstract). In: Proceedings of the [Conference Name], LNCS, vol. [Volume Number]. Springer, Heidelberg (2025). Preprint at arXiv:2501.02313

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.