From Features to Structure: Task-Aware Graph Construction for Relational and Tabular Learning with GNNs

Floris Geerts; Tamara Cucumides

REVIEW 3 major objections 6 minor 22 references

From Features to Structure: Task-Aware Graph Construction for Relational and Tabular Learning with GNNs

T0 review · 3 major / 6 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 auGraph treats graph construction for GNNs as task-aware feature selection and shows that promoting only the useful attributes into nodes beats schema-based and heuristic baselines.

desk verdict Useful idea—selective task-aware attribute promotion beats all-promote—but the GNN-gain proxy is under-specified and the empirical support is thinner than the abstract claims. read the letter →

arxiv 2506.02243 v1 pith:HDZB5AV7 submitted 2025-06-02 cs.LG

Tamara Cucumides , Floris Geerts This is my paper

classification cs.LG

keywords graphneuralnetworksrelationaldatabasestabulardataconstructionattributepromotionfeatureselectionnodeclassification

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 argues that the graph a GNN learns on should be built for the prediction task rather than fixed by the data schema. It introduces auGraph, which starts from the standard relational entity graph and repeatedly promotes non-key attributes into new attribute-value nodes, guided by scores that estimate how much each attribute helps the task. The best auGraph configuration raises F1 from 0.734 to 0.773 on synthetic relational data, from 0.930 to 0.939 on the hepatitis database, and reaches 0.985 on the mushroom table, beating the best heuristic baseline at 0.969. The practical stake is that tabular and relational GNN users get a principled alternative to hand-joining tables or blindly promoting all columns.

What carries the argument

The carrying mechanism is attribute promotion: a candidate non-key attribute $A$ is lifted into a new unary relation $R_A$ with table $T_A = \pi_A(T_i)$, so each distinct value becomes a vertex and each value vertex is linked to the row vertices that carry it. Four scoring functions rank candidates: $s_{MI}$ measures mutual information between the attribute and the label on the fully joined training table; $s_{ent}$ measures the drop in label entropy in $d$-hop neighborhoods after augmentation; $s_{dis}$ penalizes paths through the attribute that connect training nodes with different labels; and $s_{GNN}$ compares validation performance of a fixed trained GNN on the augmented versus base graph, without retraining. The iterative loop promotes the top-scoring attribute, recomputes all scores, and stops when the budget $k$ is reached or the best score falls below threshold $\tau$, so later promotions favor complementary rather than redundant attributes.

What would settle it

A decisive test: on a relational dataset with known label-relevant attributes, for each attribute compare the GNN-gain proxy — the validation score of a frozen GNN on the graph augmented with that attribute — with the validation score of a freshly retrained GNN on the same augmented graph; if these rankings disagree substantially, the selection procedure rests on an unreliable proxy and the reported gains would not be expected to transfer.

Watch

Extended reading notes

Core claim

auGraph's central claim is that attribute-level information should enter a relational entity graph selectively, as task-aware structure, rather than staying as flat features or being added wholesale. For a chosen non-key attribute $A$, the framework extracts the unary table $T_A = \pi_A(T_i)$, adds one node per distinct value, and connects each value node to every row containing that value; it then repeats this promotion with one of four scores — mutual information, entropy gain, path disagreement, or GNN gain — recomputing scores after each step until a budget $k$ or threshold $\tau$ is reached. The original schema is preserved; only structure is added. In the paper's experiments, the validation-selected auGraph configuration outperforms the relational entity graph, all-promote, random-$k$, and kNN baselines: F1 0.773 versus 0.734 on synthetic relational data, 0.939 versus 0.930 on hepatitis, and 0.985 versus 0.969 on mushroom, with the GNN-gain score best on relational tasks and mutual information best on the tabular task.

Load-bearing premise

The load-bearing premise is that a fixed model's validation score on an augmented graph, without retraining, ranks attributes in roughly the same order as a freshly retrained model's score would.

Editorial extensions

If this is right

On relational data, a GNN can beat the standard schema-derived relational entity graph without changing the GNN architecture, solely by promoting a few validation-selected attributes into nodes.
Promoting every attribute can hurt, so brute-force structural augmentation is not a safe default; task-aware selection is what delivers the gains.
For single-table data, auGraph supplies a graph construction that beats both all-promote and nearest-neighbour graphs, so tabular GNNs need not depend on feature-space geometry.
The best scoring metric differs by regime — GNN gain for relational tasks, mutual information for the tabular task — so choosing among the four scores by validation performance is part of the method rather than a tuning detail.
Because scores are recomputed after every promotion, the construction favors compact, complementary graphs, which matters when the base graph is large.

Reading between the lines

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

The fixed-model GNN-gain proxy, if it holds up, implies a cheap recipe for graph construction: train one model on the base graph, score all candidate augmentations with that frozen model, and train only once more on the final graph; a natural check is to compare this proxy against retrained performance on more datasets.
All four scoring functions need labels, so a semi-supervised setting would require pseudo-labels or representation-based scores; the paper does not address this case.
A testable extension suggested by the conclusion is to align augmentation with label agreement under 1-WL colourings; one concrete hypothesis is that label-homogeneous attribute neighborhoods predict downstream accuracy better than the current scores do.
The relational-versus-tabular split in best metric hints that model-based scores matter most where the base graph already carries structure, while statistical scores suffice where the graph is nearly empty; a broader benchmark could test whether that split is stable.

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.

Desk Editor's Note

read the letter

Colleague,

The useful new thing here is the selection loop: instead of promoting every attribute to a node, as GFS does, auGraph scores candidate attributes and iteratively promotes only the ones that look useful, recomputing scores after each promotion. That framing — graph construction as task-aware feature selection — is sensible, and the paper shows a consistent if narrow pattern: selective promotion beats all-promote and random promotion on the three benchmarks they tried. Credit where due: the code is public, the method is described clearly enough to reimplement, and the authors are honest that the expressivity link is weak.

The soft spots are real, though. Table 1 has no error bars or significance tests; three datasets is thin for a universal claim. They report the auGraph number as the best of four scoring functions chosen on validation, which inflates test performance relative to the baselines, and they don't show the other scoring functions' test numbers. The GNN-gain proxy in Section 3.2 is the biggest problem: a fixed two-layer heterogeneous GNN trained on the base REG has no trained message-passing weights for a newly added attribute-relation edge type. The paper never says how Eval_{f_theta}(G(A)) is computed — random init? copied weights? — and there's no experiment comparing GNN-gain's ranking against actually retraining. Since the relational wins in Table 1 use sGNN, the headline relational claim rests on this unvalidated proxy. I don't think the central idea is wrong; I just think the current evidence doesn't support the strong statement in the abstract.

The novelty claim is also a bit overstated. Promotion itself is from GFS [19]; the new bit is selective, iterative choice. That's a legitimate step forward, but I'd frame it as an incremental improvement rather than the 'first task-aware graph construction' claim in the conclusions.

Who gets value: this is a methods paper for people building GNN pipelines over relational or tabular data. It deserves a serious referee; a good reviewer will push for more datasets, error bars, results for all scoring metrics, and a clean explanation/validation of the GNN-gain proxy. I'd like to see the fixed-model proxy checked first, because if it doesn't hold, the relational results get shaky, but the method can still work with the other scores.

Referee Report

3 major / 6 minor

Summary. The paper proposes auGraph, a graph-construction framework for relational and tabular data. Starting from a relational entity graph (REG), it promotes selected attributes into new attribute-value nodes linked to the tuples that carry those values. Selection is guided by four scoring functions—mutual information, entropy gain, path disagreement, and a GNN-gain proxy—and is performed iteratively up to a budget k or an early-stopping threshold tau. The framework is evaluated on one synthetic relational dataset, the hepatitis multi-table dataset, and the mushroom tabular dataset, with node classification as the target task. The authors report that auGraph outperforms schema-based and heuristic baselines on these datasets.

Significance. The idea of making graph construction task-aware via attribute promotion is well-motivated, and the proposed scoring functions are simple and interpretable. The release of a public implementation supports reproducibility. If the empirical claims hold up, the method could be a useful addition to the relational and tabular GNN toolbox. However, the current evidence is too limited: the GNN-gain score, which drives the relational results, is not fully specified for unseen relation types and is never validated against retraining; the experiments lack error bars and significance tests; and the evaluation covers only three datasets. These gaps are load-bearing for the headline claim of consistent improvement, so the paper requires substantive revision rather than minor polishing.

major comments (3)

[Section 3.2 (GNN Gain)] The GNN-gain score is defined as sGNN(A)=Eval_{f_theta}(G(A)_REG) - Eval_{f_theta}(G_REG) with f_theta trained once on G_REG and kept fixed. This is not well-defined for the architecture described in Section 4.1. Promoting attribute A introduces a new relation R_A and a new edge type between attribute-value nodes and row nodes; a fixed HeteroConv/SAGEConv model has no message-passing parameters for that unseen edge type. The paper does not state how Eval_{f_theta} on G(A)_REG is computed in this situation (e.g., randomly initialized weights, weight sharing, or an extended forward pass). If untrained weights are used, the score reflects random noise rather than attribute utility. Because the relational rows of Table 1 report auGraph(top-k, sGNN), the headline relational gains depend on this unspecified computation. Please specify the exact forward pass and validate the proxy by comparing sGNN attribute rankings with rankings obtained by retraining the GNN on each candidate augmented graph.
[Table 1 / Section 4.2] The auGraph rows in Table 1 are reported as the metric that performs best on the validation set among four scoring functions. This best-of-four selection, together with the lack of standard deviations (Random-k is averaged over three runs only, and no variance is reported for any configuration), makes the 'consistently outperforms' claim hard to evaluate. Selection on the validation set is a legitimate model-selection procedure, but it should be reported as such, with per-metric test results, multiple seeds, and significance tests; otherwise the reported gains could be inflated by selection. Please add these details.
[Section 4.1 / 4.2] The evaluation is too narrow for the paper's general claims. Only one real relational dataset (hepatitis) is used, the relational evidence otherwise comes from a synthetic dataset, and the tabular evidence is a single UCI dataset (mushroom). The abstract and Section 5 claim a unified framework for relational and tabular data; that claim requires a broader benchmark suite (e.g., additional CTU datasets or RelBench) or a more modest statement of scope.

minor comments (6)

[Section 1] The word 'mechanishms' should be 'mechanisms'.
[Section 3.3] The word 'ealrly-stopping' should be 'early-stopping'.
[Figure 1 caption] The word 'Singe-table' should be 'Single-table'.
[Section 3.2 (Path Disagreement)] The normalization constant Z is not defined; please state whether it depends on the attribute A or is the same for all candidates.
[Section 3.2 (Entropy Gain)] The entropy-gain metric is undefined when a d-hop neighborhood contains no training nodes; please specify the convention used in that case.
[Section 5] The claim of being 'the first framework to pose graph construction as a task-aware feature selection problem' is too strong without a more detailed comparison to GFS [19] and other feature-selection-based graph construction methods.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: auGraph's scoring functions are selection heuristics, and reported test results come from retrained models evaluated on held-out data.

full rationale

auGraph's graph construction is an empirical preprocessing heuristic rather than a derivation whose output is equivalent to its input. The scoring functions (mutual information, entropy gain, path disagreement, and GNN-gain) are defined as measurable criteria computed from training or validation labels, and the final comparison in Table 1 reports test performance of freshly trained GNNs on the selected graph. Thus no score is definitionally equal to the reported outcome. The GNN-gain score uses a fixed pretrained model and validation labels as a proxy; this is an unvalidated modeling assumption, and the paper does not specify how a HeteroConv model handles previously unseen edge types, but it is not circular because the reported test result is obtained by retraining on the augmented graph and evaluating on held-out test data. Selection of the promotion budget k, early-stopping threshold tau, and the scoring metric uses validation performance only, not the test set, so the table entries are genuine empirical outcomes rather than fitted inputs renamed as predictions. The reference list contains no self-citations by the authors, and no load-bearing uniqueness claim is imported from prior work. The main concerns are correctness and robustness issues (fixed-model proxy validity and possible selection bias from choosing the best validation metric), which do not constitute definitional circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method has few invented entities and the main free parameters are hyperparameters selected on validation. The load-bearing assumptions are the homophily premise behind the entropy and path scores, and the fixed-model proxy in GNN-gain. These assumptions are embedded in the scoring definitions in Section 3.2 and are not separately verified.

free parameters (4)

promotion budget k = 3 in experiments, selected via validation
Controls how many attributes are promoted; chosen on validation performance, not derived from first principles.
early-stopping threshold tau = not specified numerically, selected via validation
Stops augmentation when the best available score falls below tau; the value is not reported and is chosen empirically.
GNN-gain base model f_theta = learned weights from training on the base graph
The scoring model is trained once on the relational entity graph; its weights are fitted to data and then used to evaluate augmentations.
entropy gain neighborhood depth d = not specified
The d-hop neighborhood used in the entropy gain score is not given; a user must choose this value, affecting the score.

assumptions (4)

domain assumption Promoting attribute values into nodes, with edges from tuples to their value nodes, preserves the schema and injects useful structural signal.
Section 3.1 defines this augmentation; the entire method rests on these new nodes and edges carrying predictive information for the downstream GNN.
domain assumption Label-homogeneous neighborhoods improve GNN node classification, so entropy gain and path disagreement are valid selection criteria.
Section 3.2 defines entropy gain and path disagreement based on label homogeneity; the paper itself acknowledges in Section 5 that the expressivity-performance link in GNNs is weak.
domain assumption Fixed-model validation performance on the augmented graph predicts the benefit of augmentation after retraining.
Section 3.2 defines s_GNN as the change in a frozen model's validation performance; this proxy is not validated against retrained performance in the paper.
domain assumption The fully joined training table T_train is a faithful basis for estimating mutual information even when joins create duplicate rows.
Section 3.2 computes s_MI over T_train = T_1 joins ... joins T_train; one-to-many joins can multiply rows per target tuple, but the paper does not discuss how this affects MI estimates.

invented entities (1)

Attribute-value nodes (v_a for each distinct value a of a promoted attribute A)
purpose: Connect tuples that share an attribute value, creating new graph structure that encodes feature relationships for message passing.
The paper introduces these nodes as part of auGraph; the empirical experiments suggest benefit on the tested datasets, but there is no external falsifiable handle beyond these datasets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Features to Structure: Task-Aware Graph Construction for Relational and Tabular Learning with GNNs." pith.science (2026). https://pith.science/paper/HDZB5AV7

@misc{pith2026250602243,
  author       = {Pith},
  title        = {Pith review of: From Features to Structure: Task-Aware Graph Construction for Relational and Tabular Learning with GNNs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDZB5AV7}},
  note         = {Machine review of arXiv:2506.02243}
}

read the original abstract

Tabular and relational data remain the most ubiquitous formats in real-world machine learning applications, spanning domains from finance to healthcare. Although both formats offer structured representations, they pose distinct challenges for modern deep learning methods, which typically assume flat, feature-aligned inputs. Graph Neural Networks (GNNs) have emerged as a promising solution by capturing structural dependencies within and between tables. However, existing GNN-based approaches often rely on rigid, schema-derived graphs -- such as those based on primary-foreign key links -- thereby underutilizing rich, predictive signals in non key attributes. In this work, we introduce auGraph, a unified framework for task-aware graph augmentation that applies to both tabular and relational data. auGraph enhances base graph structures by selectively promoting attributes into nodes, guided by scoring functions that quantify their relevance to the downstream prediction task. This augmentation preserves the original data schema while injecting task-relevant structural signal. Empirically, auGraph outperforms schema-based and heuristic graph construction methods by producing graphs that better support learning for relational and tabular prediction tasks.

Figures

Figures reproduced from arXiv: 2506.02243 by the authors.

**Figure 1.** Overview of auGraph. Given a tabular dataset and a task, auGraph constructs a task-aware graph through iterative attribute promotion and scoring. The resulting graph is then used by a downstream GNN to perform the learning task. 2.2 Relational Entity Graphs Given a schema (R, L) and a database 𝐷 over that schema, we define its relational entity graph as 𝐺REG = (𝑉 , 𝐸, 𝜙, 𝜏), where • the vertex set 𝑉 is defined as 𝑉 … view at source ↗

**Figure 2.** Accuracy on augmented graph with different scor [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 11 canonical work pages

[19]

Han Zhang, Quan Gan, David Wipf, and Weinan Zhang. 2023. GFS: Graph- based Feature Synthesis for Prediction over Relational Databases.arXiv preprint arXiv:2312.02037(2023). https://doi.org/10.48550/arXiv.2312.02037 arXiv:2312.02037 Presented at the VLDB 2024 Workshop on Tabular Data Anal- ysis (TaDA). 5

arXiv doi:10.48550/arxiv.2312.02037 2023
[1]

Riccardo Cappuzzo, Paolo Papotti, and Saravanan Thirumuruganathan. 2020. Creating Embeddings of Heterogeneous Relational Datasets for Data Integration Tasks. InProceedings of the 2020 ACM SIGMOD International Conference on Management of Data. Association for Computing Machinery, New York, NY, USA, 1335–1349. https://doi.org/10.1145/3318464.3389742

arXiv 2020
[2]

Milan Cvitkovic. 2020. Supervised Learning on Relational Databases with Graph Neural Networks.arXiv preprint arXiv:2002.02046(2020). https://doi.org/10. 48550/arXiv.2002.02046 arXiv:2002.02046 Presented at the ICLR 2019 Workshop on Representation Learning on Graphs and Manifolds

arXiv doi:10.48550/arxiv.2002.02046 2020
[3]

Kounianhua Du, Weinan Zhang, Ruiwen Zhou, Yangkun Wang, Xilong Zhao, Jiarui Jin, Quan Gan, Zheng Zhang, and David Wipf. 2022. Learning En- hanced Representation for Tabular Data via Neighborhood Propagation. In Advances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., 16373–16384. https://proceedings.neurips.cc/paper_files/paper...

work page 2022
[4]

Kanatsoulis, Rishi Puri, Matthias Fey, and Jure Leskovec

Vijay Prakash Dwivedi, Sri Jaladi, Yangyi Shen, Federico López, Charilaos I. Kanatsoulis, Rishi Puri, Matthias Fey, and Jure Leskovec. 2025. Relational Graph Transformer.arXiv preprint arXiv:2505.10960(2025). https://doi.org/10.48550/ arXiv.2505.10960 arXiv:2505.10960

doi:10.48550/arxiv.2505.10960 2025
[5]

Federico Errica. 2023. On Class Distributions Induced by Nearest Neigh- bor Graphs for Node Classification of Tabular Data. InAdvances in Neural Information Processing Systems, Vol. 36. Curran Associates, Inc., 28910–28940. https://proceedings.neurips.cc/paper_files/paper/2023/file/ 5c1863f711c721648387ac2ef745facb-Paper-Conference.pdf

work page 2023
[6]

Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson, Rex Ying, Jiaxuan You, and Jure Leskovec. 2024. Position: Relational Deep Learning – Graph Representation Learning on Relational Databases. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PM...

work page 2024
[7]

Matthias Fey, Vid Kocijan, Federico Lopez, Jan Eric Lenssen, and Jure Leskovec

work page

Show all 22 references

[8]

Matthias Fey and Jan Eric Lenssen. 2019. Fast Graph Representation Learning with PyTorch Geometric.arXiv preprint arXiv:1903.02428(2019). https://doi.org/ 10.48550/arXiv.1903.02428 arXiv:1903.02428

2019 doi arXiv
[9]

Xiawei Guo, Yuhan Quan, Huan Zhao, Quanming Yao, Yong Li, and Weiwei Tu
[10]

Hamilton, Zhitao Ying, and Jure Leskovec

William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive Represen- tation Learning on Large Graphs. InAdvances in Neural Information Processing Systems, Vol. 30. Curran Associates, Inc., 1024–1034. https://proceedings.neurips. cc/paper_files/paper/2017/file/5dd9db5...

2017
[11]

Kipf and Max Welling

Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. InProceedings of the 5th International Conference on Learning Representations (ICLR). OpenReview.net. https://openreview.net/ forum?id=SJU4ayYgl

2017
[12]

Cheng-Te Li, Yu-Che Tsai, Chih-Yao Chen, and Jay Chiehen Liao. 2024. Graph Neural Networks for Tabular Data Learning: A Survey with Taxonomy and Directions.arXiv preprint arXiv:2401.02143(2024). arXiv:2401.02143 https: //arxiv.org/abs/2401.02143

2024 arXiv
[13]

Jan Motl and Oliver Schulte. 2015. The CTU Prague Relational Learning Reposi- tory.arXiv preprint arXiv:1511.03086(2015). https://doi.org/10.48550/arXiv.1511. 03086 arXiv:1511.03086

2015 doi arXiv
[14]

Thomas Müller, Francesco Piccinno, Peter Shaw, Massimo Nicosia, and Yasemin Altun. 2019. Answering Conversational Questions on Structured Data with- out Logical Forms. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Internat...

2019 doi
[15]

Pu, and Renée J

Fatemeh Nargesian, Erkang Zhu, Ken Q. Pu, and Renée J. Miller. 2018. Table union search on open data.Proc. VLDB Endow.11, 7 (March 2018), 813–825. https://doi.org/10.14778/3192965.3192973

2018
[16]

Kipf, Peter Bloem, Rianne Van den Berg, Ivan Titov, and Max Welling

Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne Van den Berg, Ivan Titov, and Max Welling. 2018. Modeling Relational Data with Graph Convo- lutional Networks. InProceedings of the 15th International Conference on The Semantic Web (ESWC) (Lecture Notes in Computer Sc...

2018 doi
[17]

UCI Machine Learning Repository. 2024. Mushroom Dataset. https://doi.org/10. 24432/C5959T Accessed May 2025

2024
[18]

Antonis Vasileiou, Stefanie Jegelka, Ron Levie, and Christopher Morris
[21]

https://doi.org/10.48550/arXiv.2503.15650 arXiv:2503.15650

Survey on Generalization Theory for Graph Neural Networks.arXiv preprint arXiv:2503.15650(2025). https://doi.org/10.48550/arXiv.2503.15650 arXiv:2503.15650

2025 doi arXiv
[2021]

arXiv preprint arXiv:2108.09127(2021)

TabGNN: Multiplex Graph Neural Network for Tabular Data Prediction. arXiv preprint arXiv:2108.09127(2021). https://doi.org/10.48550/arXiv.2108.09127 arXiv:2108.09127 Presented at the KDD 2021 Workshop on Deep Learning Practice for High-Dimensional Sparse Data

2021 doi arXiv
[2025]

Kumo.ai whitepaper

KumoRFM: A Foundation Model for In-Context Learning on Rela- tional Data. Kumo.ai whitepaper. https://kumo.ai/research/kumo_relational_ foundation_model.pdf Accessed May, 2025

2025

Pith tools

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

[1] [19]

Han Zhang, Quan Gan, David Wipf, and Weinan Zhang. 2023. GFS: Graph- based Feature Synthesis for Prediction over Relational Databases.arXiv preprint arXiv:2312.02037(2023). https://doi.org/10.48550/arXiv.2312.02037 arXiv:2312.02037 Presented at the VLDB 2024 Workshop on Tabular Data Anal- ysis (TaDA). 5

arXiv doi:10.48550/arxiv.2312.02037 2023

[2] [1]

Riccardo Cappuzzo, Paolo Papotti, and Saravanan Thirumuruganathan. 2020. Creating Embeddings of Heterogeneous Relational Datasets for Data Integration Tasks. InProceedings of the 2020 ACM SIGMOD International Conference on Management of Data. Association for Computing Machinery, New York, NY, USA, 1335–1349. https://doi.org/10.1145/3318464.3389742

arXiv 2020

[3] [2]

Milan Cvitkovic. 2020. Supervised Learning on Relational Databases with Graph Neural Networks.arXiv preprint arXiv:2002.02046(2020). https://doi.org/10. 48550/arXiv.2002.02046 arXiv:2002.02046 Presented at the ICLR 2019 Workshop on Representation Learning on Graphs and Manifolds

arXiv doi:10.48550/arxiv.2002.02046 2020

[4] [3]

Kounianhua Du, Weinan Zhang, Ruiwen Zhou, Yangkun Wang, Xilong Zhao, Jiarui Jin, Quan Gan, Zheng Zhang, and David Wipf. 2022. Learning En- hanced Representation for Tabular Data via Neighborhood Propagation. In Advances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., 16373–16384. https://proceedings.neurips.cc/paper_files/paper...

work page 2022

[5] [4]

Kanatsoulis, Rishi Puri, Matthias Fey, and Jure Leskovec

Vijay Prakash Dwivedi, Sri Jaladi, Yangyi Shen, Federico López, Charilaos I. Kanatsoulis, Rishi Puri, Matthias Fey, and Jure Leskovec. 2025. Relational Graph Transformer.arXiv preprint arXiv:2505.10960(2025). https://doi.org/10.48550/ arXiv.2505.10960 arXiv:2505.10960

doi:10.48550/arxiv.2505.10960 2025

[6] [5]

Federico Errica. 2023. On Class Distributions Induced by Nearest Neigh- bor Graphs for Node Classification of Tabular Data. InAdvances in Neural Information Processing Systems, Vol. 36. Curran Associates, Inc., 28910–28940. https://proceedings.neurips.cc/paper_files/paper/2023/file/ 5c1863f711c721648387ac2ef745facb-Paper-Conference.pdf

work page 2023

[7] [6]

Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson, Rex Ying, Jiaxuan You, and Jure Leskovec. 2024. Position: Relational Deep Learning – Graph Representation Learning on Relational Databases. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PM...

work page 2024

[8] [7]

Matthias Fey, Vid Kocijan, Federico Lopez, Jan Eric Lenssen, and Jure Leskovec

work page