pith. sign in

arxiv: 2605.21475 · v1 · pith:ECIOJHQWnew · submitted 2026-05-20 · 💻 cs.LG

Is Fixing Schema Graphs Necessary? Full-Resolution Graph Structure Learning for Relational Deep Learning

Yi Huang , Qingyun Sun , Jia Li , Xingcheng Fu , Jianxin Li This is my paper

Pith reviewed 2026-05-21 05:07 UTC · model grok-4.3

classification 💻 cs.LG
keywords relational deep learninggraph structure learningrelational databasesgraph neural networkstable role modelingfunctional dependenciesschema graphsmessage passing
0
0 comments X

The pith

Relational deep learning can learn its graph structures instead of fixing them from the schema in advance.

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

The paper claims that the standard practice of building fixed full-resolution graphs from relational database schemas before applying graph neural networks is not required for good performance. It introduces a framework that turns graph construction into a learnable process by assigning roles to tables so they can serve as either nodes or edges during message passing. Role-specific passing rules then let the model adjust connections while it trains, and functional dependency rules keep the learned representations consistent at both table and entity levels. If this holds, practitioners could stop hand-designing schema graphs and instead let the data and task jointly shape the relational structure used for prediction.

Core claim

The paper establishes that full-resolution graph structure learning for relational deep learning is possible by recasting structure selection as a learnable table role modeling problem; tables thereby participate directly as nodes or edges inside message passing, role-driven passing mechanisms capture semantics, and functional dependency constraints regularize representations so that joint optimization of graph and network parameters remains semantically valid.

What carries the argument

FROG framework that formulates relational structure learning as a learnable table role modeling problem, letting tables act as nodes or edges inside message passing.

If this is right

  • Graph structure and GNN weights can be optimized together in a single end-to-end loop.
  • Role-driven message passing directly encodes how each table participates in relational semantics.
  • Representations at table and entity levels stay aligned through the added functional dependency regularization.
  • Downstream prediction accuracy on relational tasks rises relative to fixed-schema baselines.
  • Analysis of the learned roles reveals which tables function more as nodes versus edges for a given task.

Where Pith is reading between the lines

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

  • The same role-modeling idea might let other structured-data models adapt their connectivity without manual redesign.
  • Enterprise pipelines that currently require schema experts could become more automated if structure learning proves reliable.
  • Testing whether the learned graphs recover known domain rules or surface new ones would be a direct next measurement.
  • Scaling the approach to databases with thousands of tables would require checking whether role assignment stays computationally tractable.

Load-bearing premise

Functional dependency constraints are enough to keep learned representations semantically consistent when graph structure is allowed to change during training.

What would settle it

On a relational benchmark with known functional dependencies, remove the dependency constraints and measure whether accuracy falls or representations diverge across table and entity levels; a large drop would indicate the constraints are necessary.

Figures

Figures reproduced from arXiv: 2605.21475 by Jia Li, Jianxin Li, Qingyun Sun, Xingcheng Fu, Yi Huang.

Figure 1
Figure 1. Figure 1: Pipeline of Typical RDL Method tities are connected through explicit relations to describe complex systems. Because of this structured organization, such relational data is large in scale and rich in structural information, making it central to practical decision-making. Accordingly, a wide range of predictive tasks are defined on relational databases (Robinson et al., 2024; Wang et al., 2024), such as for… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the impact of graph structure rather, the RDB itself constitutes the core input, and the graph serves only as an alternative representation,which is expected to preserve two essential relational properties: full-resolution and functional dependencies. ① The full￾resolution property was first introduced by Fey et al. (2024), which requires each entity and foreign key link in the RDB to be fa… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the core concepts of RDL can be added, deleted, or modified. This is mostly achieved by modifying the adjacency matrix or Laplacian matrix of the graph (Zhu et al., 2021; Li et al., 2018). GSL makes graphs more compatible with models and downstream tasks, achieving collaborative updates of graphs and models. How￾ever, current GSL methods are inherently incompatible with RDL mapping: ① The model… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the proposed FROG method. (a) Through table-as-node and table-as-edge modeling with relation-aware message passing, the roles of tables under specific relational paradigms in the Schema Graph are learned, enabling REG structural optimization under full-resolution settings. (b) FD constraints, which regularize table-level embedding spaces and entity-level representations via discriminative model… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of table-as-node and table-as-edge Building on this result, we reinterpret the GSL objective from controlling edge existence to determining the roles of original tables, both of which govern information propaga￾tion during message passing. This enables graph structure learning from an alternative perspective. Theorem 4.1 estab￾lishes that FROG based on table-as-node and table-as-edge continues… view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of LFD The FD constraints are optimized alternating manner with GNN learning, ensuring that relational dependencies are enforced throughout the representation learning process. During GNN training, the low-rank matrix P and the scoring models are kept fixed, and the overall loss is defined as: Lmain = Ltask + LFD (17) 5. Experiment We design experiments to answer the following questions: RQ1: … view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of relational patterns learned by the model Results: The results in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: Experimental results about Table-as-Node or -Edge To further investigate the impact of modeling tables-as-node or tables-as-edge under specific relations, we design three model variants to analyze the effects of different table roles: ① Modeling all tables as nodes; ② Modeling all eligible tables as edges whenever possible; and ③ Using a random weight to balance the table-as-node and table-as-edge repre￾se… view at source ↗
Figure 9
Figure 9. Figure 9: Structure Comparison Across Different Task Types on rel-avito (Left) and rel-event (Right), Including Classification, Regression, and Link Prediction Tasks 5.6. Structural Transferability To study the transferability of learned structures across tasks, we conduct experiments on each dataset by training each task with structures learned from other tasks. We report results on rel-event and rel-avito as shown… view at source ↗
Figure 10
Figure 10. Figure 10: Hyperparameter Sensitivity Analysis of β, γ As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of rel-event and rel-avito The corresponding schema graphs are shown in [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Schema Graphs of rel-event and rel-avito F.7. Node or Edge? Analysis of Learned Graph Structures Our FROG not only produces a trained model after training, but also obtains the optimized graph structure of the Schema graph. To investigate the roles of tables under different relations, we collected the following statistics for each dataset: • The modeling roles of tables to which intermediate nodes belong … view at source ↗
Figure 13
Figure 13. Figure 13: Statistical analysis of the modeling methods of the table under different relationships 26 [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
read the original abstract

Relational prediction tasks are fundamental in many real-world applications, where data are naturally stored in relational databases (RDBs). Relational Deep Learning (RDL) addresses this problem by modeling RDBs as graphs and applying graph neural networks (GNNs) for end-to-end learning. However, the full-resolution property is commonly adopted as a design principle in graph construction for RDBs to preserve relational semantics, which leads most existing methods to rely on fixed graph structures. In this paper, we propose FROG, a Full-Resolution and Optimizable Graph Structure Learning} framework for RDL that formulates relational structure learning as a learnable table role modeling problem, allowing tables to contribute as nodes and edges in message passing. We further design role-driven message passing mechanisms to capture relational semantics, enabling joint optimization of graph structure and GNN representations. To ensure semantic consistency, we introduce functional dependency constraints that regularize representations across table and entity levels. Extensive experiments demonstrate that our method outperforms existing approaches and reveal how table roles impact downstream tasks, offering new insights into graph construction for RDL

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper proposes FROG, a Full-Resolution and Optimizable Graph Structure Learning framework for Relational Deep Learning (RDL). It formulates relational structure learning as a learnable table role modeling problem that allows tables to act as nodes and edges during message passing, introduces role-driven message passing mechanisms, and adds functional dependency constraints as regularizers across table and entity levels to preserve semantic consistency. The central claim is that this enables joint optimization of graph structure and GNN representations, outperforms prior fixed-schema approaches, and yields new insights into the impact of table roles on downstream tasks.

Significance. If the empirical results and regularization hold under scrutiny, the work would meaningfully challenge the prevailing design principle of fixed full-resolution graphs in RDL. It offers a concrete path toward end-to-end structure optimization while attempting to retain relational semantics, supported by extensive experiments that demonstrate outperformance and provide interpretable insights on table roles. These elements constitute a substantive contribution to graph construction practices for relational data.

major comments (2)
  1. [§4.3] §4.3 (Functional Dependency Regularization): the soft penalty on FD violations is presented as sufficient to maintain semantic consistency when table roles are freely optimized, yet the manuscript provides no theoretical bound or counter-example analysis showing that numerical satisfaction of the penalty necessarily restores or preserves original schema semantics, particularly for schemas containing weak or overlapping dependencies. This is load-bearing for the claim that learnable roles plus regularization jointly enable safe structure optimization.
  2. [§5.2, Table 4] §5.2 and Table 4 (Ablation studies): the performance gains attributed to the full FROG model are reported without an explicit ablation that isolates the contribution of the FD regularization term versus the role-driven message passing alone; without this isolation it remains unclear whether the constraints are actively preventing semantic drift or merely acting as a minor regularizer.
minor comments (3)
  1. [§3.1] The notation distinguishing 'table-as-node' versus 'table-as-edge' roles could be made more explicit in the method section to reduce ambiguity when readers compare against standard heterogeneous GNN formulations.
  2. [Figure 2] Figure 2 (learned graph visualizations) would benefit from side-by-side comparison with the original schema graph on at least two datasets to illustrate the magnitude of structural changes induced by optimization.
  3. A brief discussion of computational overhead introduced by the joint optimization of roles and representations would help readers assess practical deployability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and describe the revisions we will incorporate to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§4.3] §4.3 (Functional Dependency Regularization): the soft penalty on FD violations is presented as sufficient to maintain semantic consistency when table roles are freely optimized, yet the manuscript provides no theoretical bound or counter-example analysis showing that numerical satisfaction of the penalty necessarily restores or preserves original schema semantics, particularly for schemas containing weak or overlapping dependencies. This is load-bearing for the claim that learnable roles plus regularization jointly enable safe structure optimization.

    Authors: We agree that the manuscript does not supply a formal theoretical bound or exhaustive counter-example analysis demonstrating that satisfaction of the soft penalty necessarily preserves schema semantics in all cases, especially for weak or overlapping dependencies. The penalty is formulated as a differentiable term that scales with the magnitude of violation at both table and entity levels, and our experiments indicate it discourages semantically inconsistent role assignments in practice. To address the concern, we will revise §4.3 to include additional discussion of the penalty's behavior on weak dependencies, report empirical counter-examples on synthetic schemas with controlled dependency strength, and explicitly note the limitations of the soft-constraint approach as a practical rather than provably complete safeguard. revision: partial

  2. Referee: [§5.2, Table 4] §5.2 and Table 4 (Ablation studies): the performance gains attributed to the full FROG model are reported without an explicit ablation that isolates the contribution of the FD regularization term versus the role-driven message passing alone; without this isolation it remains unclear whether the constraints are actively preventing semantic drift or merely acting as a minor regularizer.

    Authors: We thank the referee for highlighting this gap. The ablations currently presented compare the complete FROG model against external baselines but do not isolate the FD regularization term from the role-driven message passing mechanism. In the revision we will add a targeted ablation in §5.2 that evaluates a variant with role-driven message passing enabled but the FD regularization disabled. The results will be incorporated into an updated Table 4 (or a new supplementary table) to clarify whether the constraints contribute meaningfully to semantic consistency beyond their effect as a general regularizer. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in FROG derivation

full rationale

The paper's core contribution is the introduction of FROG as a new framework that treats relational structure learning as a learnable table role modeling problem, with role-driven message passing and functional dependency constraints added as regularizers. These elements are presented as novel design choices validated through experiments rather than reductions of outputs to fitted inputs or prior self-citations. No equations or claims in the abstract or description reduce a prediction to a definition by construction, nor do they rely on load-bearing self-citations or smuggled ansatzes. The derivation chain remains self-contained, with semantic consistency enforced via explicit constraints and empirical demonstration instead of tautological equivalence to the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are detailed in the provided text.

pith-pipeline@v0.9.0 · 5728 in / 1141 out tokens · 40080 ms · 2026-05-21T05:07:02.070529+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

68 extracted references · 68 canonical work pages · 2 internal anchors

  1. [1]

    NeurIPS , year=

    Relbench: A benchmark for deep learning on relational databases , author=. NeurIPS , year=

    work page

  2. [3]

    ICML , year=

    RelGNN: Composite Message Passing for Relational Deep Learning , author=. ICML , year=

    work page

  3. [4]

    NeurIPS , year=

    Universal graph convolutional networks , author=. NeurIPS , year=

    work page

  4. [5]

    ICML , year=

    Position: Relational deep learning-graph representation learning on relational databases , author=. ICML , year=

    work page

  5. [6]

    NeurIPS , year=

    4DBInfer: A 4d benchmarking toolbox for graph-centric predictive modeling on RDBs , author=. NeurIPS , year=

    work page

  6. [9]

    ICML , year=

    Griffin: Towards a Graph-Centric Relational Database Foundation Model , author=. ICML , year=

    work page

  7. [10]

    NeurIPS , year=

    A Pre-training Framework for Relational Data with Information-theoretic Principles , author=. NeurIPS , year=

    work page

  8. [11]

    AAAI , year=

    Graph structure learning with variational information bottleneck , author=. AAAI , year=

    work page

  9. [12]

    AAAI , year=

    DG-Mamba: Robust and efficient dynamic graph structure learning with selective state space models , author=. AAAI , year=

    work page

  10. [13]

    Proceedings of the ACM Web Conference 2022 , pages=

    Towards unsupervised deep graph structure learning , author=. Proceedings of the ACM Web Conference 2022 , pages=

    work page 2022

  11. [14]

    AAAI , year=

    Identity-aware graph neural networks , author=. AAAI , year=

    work page

  12. [15]

    NeurIPS , volume=

    Inductive representation learning on large graphs , author=. NeurIPS , volume=

    work page

  13. [16]

    NeurIPS , volume=

    Lightgbm: A highly efficient gradient boosting decision tree , author=. NeurIPS , volume=

    work page

  14. [18]

    SIGKDD , pages =

    Tianqi Chen and Carlos Guestrin , title =. SIGKDD , pages =

    work page

  15. [19]

    ICML , year=

    Neural message passing for quantum chemistry , author=. ICML , year=

    work page

  16. [20]

    IEEE transactions on neural networks and learning systems , volume=

    A comprehensive survey on graph neural networks , author=. IEEE transactions on neural networks and learning systems , volume=. 2020 , publisher=

    work page 2020

  17. [21]

    , author=

    A method of comparing the areas under receiver operating characteristic curves derived from the same cases. , author=. Radiology , volume=

    work page

  18. [22]

    IFIP congress , volume=

    Dependency structures of data base relationships , author=. IFIP congress , volume=

    work page

  19. [23]

    International journal of forecasting , volume=

    Forecasting seasonals and trends by exponentially weighted moving averages , author=. International journal of forecasting , volume=. 2004 , publisher=

    work page 2004

  20. [24]

    SIGKDD , pages=

    Graph structure learning for robust graph neural networks , author=. SIGKDD , pages=

    work page

  21. [26]

    AAAI , year=

    Adaptive graph convolutional neural networks , author=. AAAI , year=

    work page

  22. [27]

    SIGKDD , pages=

    Relational deep learning: Challenges, foundations and next-generation architectures , author=. SIGKDD , pages=

    work page

  23. [28]

    Artificial Intelligence Review , volume=

    Heterogeneous graph neural networks analysis: a survey of techniques, evaluations and applications , author=. Artificial Intelligence Review , volume=. 2023 , publisher=

    work page 2023

  24. [29]

    NeurIPS , year=

    Attention is all you need , author=. NeurIPS , year=

    work page

  25. [30]

    NeurIPS 2024 Third Table Representation Learning Workshop , year=

    PyTorch Frame: A Modular Framework for Multi-Modal Tabular Learning , author=. NeurIPS 2024 Third Table Representation Learning Workshop , year=

    work page 2024

  26. [31]

    European semantic web conference , pages=

    Modeling relational data with graph convolutional networks , author=. European semantic web conference , pages=. 2018 , organization=

    work page 2018

  27. [33]

    Patterns , volume=

    Leakage and the reproducibility crisis in machine-learning-based science , author=. Patterns , volume=. 2023 , publisher=

    work page 2023

  28. [34]

    ICML , year=

    Do We Really Need Message Passing in Brain Network Modeling? , author=. ICML , year=

    work page

  29. [35]

    2016 , publisher=

    Relational database design and implementation , author=. 2016 , publisher=

    work page 2016

  30. [36]

    Data Mining and Knowledge Discovery , volume=

    Hierarchical message-passing graph neural networks , author=. Data Mining and Knowledge Discovery , volume=

    work page

  31. [37]

    EMNLP , pages=

    Glove: Global vectors for word representation , author=. EMNLP , pages=

    work page

  32. [38]

    Jiaming Zhuo and Can Cui and Kun Fu and Bingxin Niu and Dongxiao He and Yuanfang Guo and Zhen Wang and Chuan Wang and Xiaochun Cao and Liang Yang , title =

    work page

  33. [39]

    Armstrong, W. W. Dependency structures of data base relationships. In IFIP congress, volume 74, pp.\ 580--583, 1974

    work page 1974

  34. [40]

    Heterogeneous graph neural networks analysis: a survey of techniques, evaluations and applications

    Bing, R., Yuan, G., Zhu, M., Meng, F., Ma, H., and Qiao, S. Heterogeneous graph neural networks analysis: a survey of techniques, evaluations and applications. Artificial Intelligence Review, 56 0 (8): 0 8003--8042, 2023

    work page 2023

  35. [41]

    and Guestrin, C

    Chen, T. and Guestrin, C. Xgboost: A scalable tree boosting system. In SIGKDD, pp.\ 785--794. ACM , 2016

    work page 2016

  36. [42]

    Relgnn: Composite message passing for relational deep learning

    Chen, T., Kanatsoulis, C., and Leskovec, J. Relgnn: Composite message passing for relational deep learning. In ICML, 2025

    work page 2025

  37. [43]

    Rdb2g-bench: A comprehensive benchmark for automatic graph modeling of relational databases

    Choi, D., Kim, S., Kim, J., Kim, K., Lee, G., Kang, S., Kim, M., and Shin, K. Rdb2g-bench: A comprehensive benchmark for automatic graph modeling of relational databases. arXiv preprint arXiv:2506.01360, 2025

    work page arXiv 2025

  38. [44]

    P., Jaladi, S., Shen, Y., L \'o pez, F., Kanatsoulis, C

    Dwivedi, V. P., Jaladi, S., Shen, Y., L \'o pez, F., Kanatsoulis, C. I., Puri, R., Fey, M., and Leskovec, J. Relational graph transformer. arXiv preprint arXiv:2505.10960, 2025 a

    work page arXiv 2025

  39. [45]

    P., Kanatsoulis, C., Huang, S., and Leskovec, J

    Dwivedi, V. P., Kanatsoulis, C., Huang, S., and Leskovec, J. Relational deep learning: Challenges, foundations and next-generation architectures. In SIGKDD, pp.\ 5999--6009, 2025 b

    work page 2025

  40. [46]

    Fast Graph Representation Learning with PyTorch Geometric

    Fey, M. and Lenssen, J. E. Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  41. [47]

    E., Ranjan, R., Robinson, J., Ying, R., You, J., and Leskovec, J

    Fey, M., Hu, W., Huang, K., Lenssen, J. E., Ranjan, R., Robinson, J., Ying, R., You, J., and Leskovec, J. Position: Relational deep learning-graph representation learning on relational databases. In ICML, 2024

    work page 2024

  42. [48]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In ICML, 2017

    work page 2017

  43. [49]

    Inductive representation learning on large graphs

    Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. NeurIPS, 30, 2017

    work page 2017

  44. [50]

    Hanley, J. A. and McNeil, B. J. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology, 148 0 (3): 0 839--843, 1983

    work page 1983

  45. [51]

    Harrington, J. L. Relational database design and implementation. Morgan Kaufmann, 2016

    work page 2016

  46. [52]

    Holt, C. C. Forecasting seasonals and trends by exponentially weighted moving averages. International journal of forecasting, 20 0 (1): 0 5--10, 2004

    work page 2004

  47. [53]

    Pytorch frame: A modular framework for multi-modal tabular learning

    Hu, W., Yuan, Y., Zhang, Z., Nitta, A., Cao, K., Kocijan, V., Sunil, J., Leskovec, J., and Fey, M. Pytorch frame: A modular framework for multi-modal tabular learning. In NeurIPS 2024 Third Table Representation Learning Workshop, 2024

    work page 2024

  48. [54]

    Graph structure learning for robust graph neural networks

    Jin, W., Ma, Y., Liu, X., Tang, X., Wang, S., and Tang, J. Graph structure learning for robust graph neural networks. In SIGKDD, pp.\ 66--74, 2020

    work page 2020

  49. [55]

    Kaggle Data Science & Machine Learning Survey , 2022

    Kaggle . Kaggle Data Science & Machine Learning Survey , 2022. Available: https://www.kaggle.com/code/paultimo thymooney/kaggle-survey-2022-all-results/notebook

    work page 2022

  50. [56]

    and Narayanan, A

    Kapoor, S. and Narayanan, A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4 0 (9), 2023

    work page 2023

  51. [57]

    Lightgbm: A highly efficient gradient boosting decision tree

    Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., and Liu, T.-Y. Lightgbm: A highly efficient gradient boosting decision tree. NeurIPS, 30, 2017

    work page 2017

  52. [58]

    Lachi, D., Mohammadi, M., Meyer, J., Arora, V., Palczewski, T., and Dyer, E. L. Integrating temporal and structural context in graph transformers for relational deep learning. arXiv preprint arXiv:2511.04557, 2025

    work page arXiv 2025

  53. [59]

    Adaptive graph convolutional neural networks

    Li, R., Wang, S., Zhu, F., and Huang, J. Adaptive graph convolutional neural networks. In AAAI, 2018

    work page 2018

  54. [60]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  55. [61]

    Pennington, J., Socher, R., and Manning, C. D. Glove: Global vectors for word representation. In EMNLP, pp.\ 1532--1543, 2014

    work page 2014

  56. [62]

    E., Yuan, Y., Zhang, Z., et al

    Robinson, J., Ranjan, R., Hu, W., Huang, K., Han, J., Dobles, A., Fey, M., Lenssen, J. E., Yuan, Y., Zhang, Z., et al. Relbench: A benchmark for deep learning on relational databases. NeurIPS, 2024

    work page 2024

  57. [63]

    Sun, Q., Li, J., Peng, H., Wu, J., Fu, X., Ji, C., and Yu, P. S. Graph structure learning with variational information bottleneck. In AAAI, 2022

    work page 2022

  58. [64]

    A pre-training framework for relational data with information-theoretic principles

    Truong, Q., Chen, Z., Ju, M., Zhao, T., Shah, N., and Tang, J. A pre-training framework for relational data with information-theoretic principles. In NeurIPS, 2025

    work page 2025

  59. [65]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. NeurIPS, 2017

    work page 2017

  60. [66]

    4dbinfer: A 4d benchmarking toolbox for graph-centric predictive modeling on rdbs

    Wang, M., Gan, Q., Wipf, D., Zhang, Z., Faloutsos, C., Zhang, W., Zhang, M., Cai, Z., Li, J., Mao, Z., et al. 4dbinfer: A 4d benchmarking toolbox for graph-centric predictive modeling on rdbs. NeurIPS, 2024

    work page 2024

  61. [67]

    Griffin: Towards a graph-centric relational database foundation model

    Wang, Y., Wang, X., Gan, Q., Wang, M., Yang, Q., Wipf, D., and Zhang, M. Griffin: Towards a graph-centric relational database foundation model. In ICML, 2025

    work page 2025

  62. [68]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Yu, P. S. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32 0 (1): 0 4--24, 2020

    work page 2020

  63. [69]

    Do we really need message passing in brain network modeling? In ICML, 2025

    Yang, L., Liu, Y., Zhuo, J., Jin, D., Wang, C., Wang, Z., and Cao, X. Do we really need message passing in brain network modeling? In ICML, 2025

    work page 2025

  64. [70]

    M., Ying, R., and Leskovec, J

    You, J., Gomes-Selman, J. M., Ying, R., and Leskovec, J. Identity-aware graph neural networks. In AAAI, 2021

    work page 2021

  65. [71]

    Dg-mamba: Robust and efficient dynamic graph structure learning with selective state space models

    Yuan, H., Sun, Q., Wang, Z., Fu, X., Ji, C., Wang, Y., Jin, B., and Li, J. Dg-mamba: Robust and efficient dynamic graph structure learning with selective state space models. In AAAI, 2025

    work page 2025

  66. [72]

    Hierarchical message-passing graph neural networks

    Zhong, Z., Li, C.-T., and Pang, J. Hierarchical message-passing graph neural networks. Data Mining and Knowledge Discovery, 37 0 (1): 0 381--408, 2023

    work page 2023

  67. [73]

    Deep graph structure learning for robust representations: A survey

    Zhu, Y., Xu, W., Zhang, J., Liu, Q., Wu, S., and Wang, L. Deep graph structure learning for robust representations: A survey. arXiv preprint arXiv:2103.03036, 14: 0 1--1, 2021

    work page arXiv 2021

  68. [74]

    Propagation is all you need: A new framework for representation learning and classifier training on graphs

    Zhuo, J., Cui, C., Fu, K., Niu, B., He, D., Guo, Y., Wang, Z., Wang, C., Cao, X., and Yang, L. Propagation is all you need: A new framework for representation learning and classifier training on graphs. In MM , pp.\ 481--489, 2023

    work page 2023