Pith. sign in

REVIEW 4 major objections 3 minor 47 references

Graph-Based Feature Augmentation for Predictive Tasks on Relational Datasets

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ReCoGNN, a two-stage graph framework, automates feature augmentation and beats filter, wrapper, and embedding baselines on ten datasets.

desk verdict ReCoGNN is a plausible two-stage feature-augmentation idea, but the label-propagation step is under-specified and the experiments don't yet back the 'consistently outperforms' claim. read the letter →

arxiv 2508.20986 v1 pith:HQS3GWJ4 submitted 2025-08-28 cs.DB cs.LG

Lianpeng Qiao , Ziqi Cao , Kaiyu Feng , Ye Yuan , Guoren Wang This is my paper
classification cs.DBcs.LG
keywords featureaugmentationrelationaldatabaseheterogeneousgraphattentionnetworkneuralattributegroupingselectionAutoML
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 feature augmentation over relational tables — pulling in columns from other, connected tables to improve a prediction — can be automated end-to-end rather than left to manual join design and feature selection. Its framework, ReCoGNN, works in two learned stages: a graph attention network discovers which attribute combinations within each auxiliary table matter for the target, splitting tables into focused sub-tables; then a weighted heterogeneous graph with join-based and similarity-based edges lets message passing enrich base-table embeddings while learned edge weights filter noisy or irrelevant attributes. The authors report that this pipeline beats filter-based, wrapper-based, tree-embedding, and prior end-to-end relational baselines on ten classification and regression datasets. If the claim holds, AutoML pipelines could accept a database schema as input and receive ready-to-use enriched features, turning augmentation from manual engineering into a learned step.

What carries the argument

The engine is the learned attribute-relation signal plus the weighted heterogeneous graph built from it. Stage One builds a complete graph over the non-key attributes of each task-relevant auxiliary tuple; a shared-parameter GAT scores every attribute pair, the scores are pooled across tuples, a threshold keeps the significant pairs, and maximal cliques of the significant-edge graph become the sub-table schemas. Stage Two turns base tuples and sub-table tuples into nodes of a heterogeneous graph, with explicit edges for joinable tuples and implicit edges for top-K similar base tuples; heterogeneous message passing with learned edge weights both selects attributes (high-weight edges mark task

What would settle it

Reshuffle the foreign keys in an auxiliary table so join paths become random while schema and value distributions stay intact, then rerun ReCoGNN on that table. If the reported gains over the base-table baseline persist, the label-propagation assumption is not load-bearing; if performance collapses toward the Base baseline, the assumption is the crux. A companion check: the paper's own ablation replaces the GAT-derived sub-tables with random attribute groupings of the same size and shows a drop, so the decisive question is whether that drop matches the noise injected by the shuffled joins.

Watch

Extended reading notes

Core claim

ReCoGNN claims that feature augmentation over relational tables reduces to two learned operations. First, a shared-parameter Graph Attention Network on complete per-tuple graphs learns which attribute combinations in each auxiliary table predict the target; summing, normalizing, and thresholding those attention weights yields a significant-edge graph whose maximal cliques become sub-tables. Second, a weighted heterogeneous graph of join and similarity edges lets message-passing GNNs learn edge weights that filter noise while enriching base-table embeddings. On ten classification and regression datasets this design beats filter, wrapper, and embedding baselines.

Load-bearing premise

The first stage is supervised entirely by labels copied onto auxiliary tuples through join paths: Section 4.1 assumes that whenever a join path links an auxiliary tuple to a base tuple, the auxiliary tuple can be treated as carrying that base tuple's target label. If joins are many-to-many, noisy, or semantically weak, the learned attention weights and the sub-tables built from them inherit that noise, and the whole pipeline degrades.

Editorial extensions

If this is right

  • If the reported gains hold, AutoML systems can treat feature augmentation as a learned step: the framework consumes the schema and outputs ready base-table embeddings, removing manual join-path design.
  • Whole-table joins become unnecessary: the attention-derived sub-tables isolate the informative attribute combinations, which is exactly where the 'All' baseline degrades under noise and data drift.
  • The edge weights learned during message passing serve double duty as a feature-selection signal, so augmentation and selection happen in a single pass rather than separate join-then-rank steps.
  • Because the same pipeline feeds both classification and regression heads, the framework is task-agnostic across the two dominant predictive settings.
  • The inductive design means newly arriving tuples can be embedded and scored without rebuilding or retraining on the full heterogeneous graph.

Reading between the lines

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

  • The label-propagation premise makes meta-path choice consequential: on schemas with many-to-many or semantically loose joins, a learnable meta-path scorer trained by the downstream loss would likely be more robust than the fixed greedy path score — a testable variant the paper does not explore.
  • Maximal-clique extraction and the threshold ℓ interact with table width; the paper's own sensitivity sweep shows a performance peak in ℓ, suggesting a per-table or learned threshold would help on wide auxiliary tables.
  • Implicit similarity edges are only placed between base-table tuples; extending them to auxiliary sub-table tuples could carry supervision further, though it risks label leakage and would need edge-type constraints.
  • Scope note for readers: the experimental section caps each trial at 18 hours with automatic termination, so reported baseline runs should be checked for completeness before comparing runtime or convergence claims.
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 / 3 minor

Summary. The paper proposes ReCoGNN, a two-stage feature augmentation framework for predictive tasks over relational tables. In Stage One, it identifies task-relevant tuples via meta-path joins, constructs a complete graph over each tuple's attributes, trains a shared GAT to predict propagated target labels, and groups attributes into sub-tables from thresholded cumulative attention matrices and maximal cliques. In Stage Two, it constructs a heterogeneous graph with explicit PK-FK join edges and implicit similarity edges among base-table tuples, and runs GNN message passing to refine base-table node embeddings for a downstream prediction head. The paper reports experiments on ten datasets and claims consistent superiority over filter, wrapper, and embedding baselines.

Significance. ReCoGNN targets an important AutoML problem: automatically finding useful augmenting features across relational tables. The design is plausible and has a clear novel ingredient: partitioning auxiliary tables into attribute sub-tables based on GAT attention, rather than treating table tuples monolithically. The paper also conducts ablations on edge weights and similarity edges. However, the empirical evidence is currently conditional: the evaluation section lacks a described validation protocol and error bars; the label-propagation step in Stage One is underspecified for one-to-many joins; and the reported results contradict the 'consistently outperforms' claim on Olist. No code or artifact is provided, so the machine-checked/reproducibility strengths are absent.

major comments (4)
  1. [§4.1, Eqs. (3)–(6)] Label propagation is underspecified. Section 4.1 states that an auxiliary tuple t_ij is considered relevant to base tuple t_0k and 'carrying the corresponding target label' whenever a join path links them. In the paper's own schemas, joins are one-to-many: a single Users row can join to multiple Event-Interest rows with different interest values. No rule is given for resolving multiple labels for one auxiliary tuple. This is load-bearing because these labels supervise the GAT (Eqs. 3–5) and the cumulative attention matrix A_sum (Eq. 6) that feeds threshold ℓ and maximal-clique sub-table extraction. Until the disambiguation rule is specified, Stage One is not well-defined. Figure 3's example is also internally inconsistent: the text says the tuple with user=1 obtains a label by linking to an Event-Interest tuple with user=4.
  2. [§6.1, §6.2, Tables 2 and 3] The experimental protocol is not described. Tables 2 and 3 report point estimates with no error bars, no repeated trials, and no standard deviations. The implementation paragraph mentions random sampling for Olist, Event, PED, and Loyal, but does not describe how train/validation/test splits were formed. Because Stage One selects sub-tables using target labels and Stage Two trains a GNN on base-table labels, an independent split is essential to avoid supervised feature selection evaluated on the training signal. Please state the split protocol, report variance over repetitions, and provide significance tests. Several cells are also missing (e.g., BE on Event/Event-Not in Table 2; XGBoost/RF on IMDB in Table 3), weakening the comparison.
  3. [Abstract, §6.2, Table 2] The claim that ReCoGNN 'consistently outperforms' existing methods is contradicted by the paper's own Olist results. In Table 2, LightGBM achieves Accuracy 0.9130 vs. ReCoGNN's 0.9084 and AUC-ROC 0.8984 vs. 0.8855, and the text states that LightGBM achieved the best results in both metrics on Olist. The central claim needs to be qualified by dataset, or supported by paired statistical tests, before it can be accepted as stated.
  4. [§6.1 baseline list] The baseline set omits any graph-based relational learning method such as [13] or [14], even though ReCoGNN builds directly on those ideas and cites [14] for its encoder. Since the paper's central empirical claim is superiority over existing methods, the absence of an end-to-end relational GNN baseline makes the comparison incomplete. Please add RDL or a comparable relational GNN method, or justify why it cannot be applied.
minor comments (3)
  1. [§6.1, Table 3] The table header says 'Regress Task' instead of 'Regression Task', and the text swaps the definitions of MAE and MSE (it says 'Mean Squared Error (MAE) and Mean Absolute Error (MSE)').
  2. [§6.2] The reported runtimes for XGBoost, LightGBM, and ReCoGNN are not accompanied by hardware details, number of repetitions, or dataset size at the time of timing; as reported, they are not directly comparable.
  3. [§4.1] The meta-path scoring function introduces tunable weights α and β, and the 1:n penalty weights w_i, but the paper does not state how these are set in the experiments. Please report the chosen values or the tuning procedure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the label-supervised feature-selection pipeline is self-contained; the flagged issues are experimental validity concerns, not definitional reductions.

full rationale

The paper's derivation chain is not circular. Stage One (§4.2) trains a GAT to predict labels propagated from the base table to join-reachable auxiliary tuples (§4.1); the learned attention matrix A_sum (Eq. 6) is then thresholded and used to form sub-tables. This is supervised feature/structure selection, not a self-referential definition: the target variable remains the base table's label, and the auxiliary tuple labels are derived from that target via schema-defined join paths. A_sum is a function of the trained GAT, but the paper explicitly uses the target to supervise the GAT, so the attribute relationships are not claimed to be independent of the label. Stage Two (§5) trains a heterogeneous GNN to predict the same base-table labels, which is standard supervised learning. There is no load-bearing self-citation or imported uniqueness theorem; the cited graph/tabular works are external. The concerns flagged by a skeptical reader—ambiguous multi-label joins for a single auxiliary tuple, and the absence of an explicit train/test split—are correctness/reproducibility issues, not circular reductions. In particular, the paper never defines the final prediction as the same quantity used to select features in a way that would make the reported superiority true by construction. Therefore score 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on two supervised learning stages whose supervision is the base table target propagated through joins. No external benchmark, formal proof, or released artifact is provided; every step is evaluated with in-paper datasets. The dominant free parameters (ℓ, α/β, top-K, coreset size) are set by hand and not reported in full.

free parameters (6)
  • ℓ (significant-edge threshold) = 0.8 (default)
    Used to convert normalized GAT attention sums into a significant-edge set before clique extraction; author-set default, sensitivity analysis in Fig. 6 shows performance varies with ℓ.
  • α and β (meta-path score weights) = not reported
    S_path = αS_L + βS_N, where α/β balance path length and 1:n join penalty; values not given in paper.
  • w_i (per-edge 1:n join risk weights) = not reported
    Used in S_N penalty for one-to-many joins; no values or fitting procedure disclosed.
  • Top-K or θ for implicit similarity edges = not reported
    Base-table similarity edges are added by threshold θ or top-K; neither value is reported.
  • Coreset size/selection criteria = not reported
    Coreset construction 'based on criteria such as label distribution' is not specified; affects which tuples train the GAT.
  • Embedding/projection dimensions (d_num, d_cat, d_text, d_out) = not reported
    Encoder and projection dimensions are not specified, affecting all learned representations.
assumptions (6)
  • domain assumption Auxiliary tuples linked via join paths inherit the base table target label.
    Section 4.1: task-relevant tuples are identified by join-path linkage and then used as supervised examples for GAT training.
  • domain assumption GAT attention weights optimized for target prediction indicate predictive attribute relationships.
    Section 4.2.2, Eq. 3-5: edge attention is used as the importance signal for grouping attributes into sub-tables.
  • ad hoc to paper Maximal cliques of high-attention edges form semantically coherent sub-tables.
    Section 4.3.2: sub-table extraction is defined as maximal cliques of significant edges; no theoretical or empirical justification is given beyond the clique definition.
  • domain assumption Similarity edges within the base table improve label propagation and node embeddings.
    Section 5.1 'Implicit Edges': base-table tuples are connected by feature similarity so GNN can spread information; no evidence beyond ablation.
  • domain assumption Coreset sampling preserves the information needed for attribute relationship mining.
    Section 4.1 'Coreset': sampling representative tuples from each table is claimed to maintain structural details; the preservation property is not proven or measured.
  • domain assumption Learned GNN edge weights can be interpreted as feature importance for filtering attributes.
    Section 5.2: the paper equates high-weight edges with important features, using this to justify feature selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-Based Feature Augmentation for Predictive Tasks on Relational Datasets." pith.science (2026). https://pith.science/paper/HQS3GWJ4

@misc{pith2026250820986,
  author       = {Pith},
  title        = {Pith review of: Graph-Based Feature Augmentation for Predictive Tasks on Relational Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HQS3GWJ4}},
  note         = {Machine review of arXiv:2508.20986}
}
read the original abstract

Data has become a foundational asset driving innovation across domains such as finance, healthcare, and e-commerce. In these areas, predictive modeling over relational tables is commonly employed, with increasing emphasis on reducing manual effort through automated machine learning (AutoML) techniques. This raises an interesting question: can feature augmentation itself be automated and identify and utilize task-related relational signals? To address this challenge, we propose an end-to-end automated feature augmentation framework, ReCoGNN, which enhances initial datasets using features extracted from multiple relational tables to support predictive tasks. ReCoGNN first captures semantic dependencies within each table by modeling intra-table attribute relationships, enabling it to partition tables into structured, semantically coherent segments. It then constructs a heterogeneous weighted graph that represents inter-row relationships across all segments. Finally, ReCoGNN leverages message-passing graph neural networks to propagate information through the graph, guiding feature selection and augmenting the original dataset. Extensive experiments conducted on ten real-life and synthetic datasets demonstrate that ReCoGNN consistently outperforms existing methods on both classification and regression tasks.

Figures

Figures reproduced from arXiv: 2508.20986 by the authors.

Figure 1
Figure 1. Event-Recommendation Relational Tables invitation status and confirmed attendance respectively. These in￾dicators reflect event popularity and may influence user interest. Such attributes can correlate with our prediction target, providing valuable auxiliary signals. Identifying attribute correlations across auxiliary tables enhances predictive performance. However, detect￾ing these relationships in complex database… view at source ↗
Figure 2
Figure 2. ReCoGNN Framework 4.1 Identifying Task-Relevant Tuples Recall that our goal is to uncover meaningful combinations of attributes within each table that contribute to predicting the tar￾get attribute’s value. To achieve this, we analyze the relationships between attribute values within individual tuples of the table. Iden￾tifying task-relevant tuples is non-trivial. The natural intuition suggests that a tuple is task-… view at source ↗
Figure 4
Figure 4. Modeling Attribute Relationship with GAT [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Scenarios for Extracting sub-tables from a Table [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Performance of Different ℓ on MovieLens Dataset [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Ablation Studies of Attributes Relationship Discov [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Ablation Studies on Edge Weights and Added Simi [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 28 canonical work pages

  1. [13]

    Milan Cvitkovic. 2020. Supervised learning on relational databases with graph neural networks. arXiv preprint arXiv:2002.02046 (2020)

    arXiv 2020

  2. [14]

    Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson, Rex Ying, Jiaxuan You, and Jure Leskovec. 2023. Relational deep learning: Graph representation learning on relational databases. arXiv preprint arXiv:2312.04615 (2023)

    arXiv 2023

  3. [1]

    https://sourceforge.net/projects/proper/files/Datasets/raw/

    [n.d.]. https://sourceforge.net/projects/proper/files/Datasets/raw/

    work page

  4. [2]

    Brazilian E-Commerce Public Dataset by Olist

    [n.d.]. Brazilian E-Commerce Public Dataset by Olist. https://www.kaggle.com/ datasets/olistbr/brazilian-ecommerce/

    work page

  5. [3]

    Elo Merchant Category Recommendation

    [n.d.]. Elo Merchant Category Recommendation. https://www.kaggle.com/c/elo- merchant-category-recommendation

    work page

  6. [4]

    Event Recommendation Engine Challenge

    [n.d.]. Event Recommendation Engine Challenge. https://www.kaggle.com/ competitions/event-recommendation-engine-challenge

    work page

  7. [5]

    Formula One

    [n.d.]. Formula One. https://ergast.com/mrd/

    work page

  8. [6]

    MovieLens 1M Dataset

    [n.d.]. MovieLens 1M Dataset. https://grouplens.org/datasets/movielens/1m/

    work page

Show all 47 references
  1. [7]

    Predicting Excitement at DonorsChoose

    [n.d.]. Predicting Excitement at DonorsChoose. https://www.kaggle.com/c/kdd- cup-2014-predicting-excitement-at-donors-choose

    2014

  2. [8]

    Hendrik Blockeel, Sašo Džeroski, Boris Kompare, Stefan Kramer, Bernhard Pfahringer, and WIM VAN LAER. 2004. Experiments in predicting biodegrad- ability. Applied Artificial Intelligence 18, 2 (2004), 157–181

    2004

  3. [9]

    Leo Breiman. 2001. Random forests. Machine learning 45, 1 (2001), 5–32

    2001

  4. [10]

    Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining . 785–794

    2016

  5. [11]

    Nadiia Chepurko, Ryan Marcus, Emanuel Zgraggen, Raul Castro Fernandez, Tim Kraska, and David Karger. 2020. ARDA: automatic relational data augmentation for machine learning. arXiv preprint arXiv:2003.09758 (2020)

    2020 arXiv

  6. [12]

    Jillian M Clements, Di Xu, Nooshin Yousefi, and Dmitry Efimov. 2020. Sequential deep learning for credit risk monitoring with tabular financial data.arXiv preprint arXiv:2012.15330 (2020)

    2020 arXiv

  7. [15]

    Sainyam Galhotra, Yue Gong, and Raul Castro Fernandez. 2023. Metam: Goal- oriented data discovery. In 2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2780–2793

    2023

  8. [16]

    Yury Gorishniy, Ivan Rubachev, and Artem Babenko. 2022. On embeddings for numerical features in tabular deep learning. Advances in Neural Information Processing Systems 35 (2022), 24991–25004

    2022

  9. [17]

    Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. 2021. Revisiting deep learning models for tabular data. Advances in Neural Information Processing Systems 34 (2021), 18932–18943

    2021

  10. [18]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. arXiv preprint arXiv:1703.04247 (2017)

    2017 arXiv

  11. [19]

    Isabelle Guyon and André Elisseeff. 2003. An introduction to variable and feature selection. Journal of machine learning research 3, Mar (2003), 1157–1182

    2003

  12. [20]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  13. [21]

    Weihua Hu, Yiwen Yuan, Zecheng Zhang, Akihiro Nitta, Kaidi Cao, Vid Kocijan, Jure Leskovec, and Matthias Fey. 2024. PyTorch Frame: A Modular Framework for Multi-Modal Tabular Learning. arXiv preprint arXiv:2404.00776 (2024)

    2024 arXiv

  14. [22]

    Andra Ionescu, Kiril Vasilev, Florena Buse, Rihan Hai, and Asterios Katsifodimos

  15. [23]

    Yuchin Juan, Yong Zhuang, Wei-Sheng Chin, and Chih-Jen Lin. 2016. Field- aware factorization machines for CTR prediction. In Proceedings of the 10th ACM conference on recommender systems . 43–50

    2016

  16. [24]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems 30 (2017)

    2017

  17. [25]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  18. [26]

    Arun Kumar, Jeffrey Naughton, Jignesh M Patel, and Xiaojin Zhu. 2016. To join or not to join? thinking twice about joins before feature selection. In Proceedings of the 2016 International Conference on Management of Data . 19–34

    2016

  19. [27]

    Viktor Leis, Andrey Gubichev, Atanas Mirchev, Peter Boncz, Alfons Kemper, and Thomas Neumann. 2015. How good are query optimizers, really? Proceedings of the VLDB Endowment 9, 3 (2015), 204–215

    2015

  20. [28]

    Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. 2019. Fi-gnn: Modeling feature interactions via graph neural networks for ctr prediction. In Proceedings of the 28th ACM international conference on information and knowledge management. 539–548

    2019

  21. [29]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data ...

    2018

  22. [30]

    Jiaming Liang, Chuan Lei, Xiao Qin, Jiani Zhang, Asterios Katsifodimos, Christos Faloutsos, and Huzefa Rangwala. 2025. FeatPilot: Automatic feature augmenta- tion on tabular data. (2025)

    2025

  23. [31]

    Jiabin Liu, Chengliang Chai, Yuyu Luo, Yin Lou, Jianhua Feng, and Nan Tang

  24. [32]

    N Reimers. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT- Networks. arXiv preprint arXiv:1908.10084 (2019)

    2019 arXiv

  25. [33]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International conference on data mining. IEEE, 995–1000

    2010

  26. [34]

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolu- tional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, pro...

    2018

  27. [35]

    Vraj Shah, Arun Kumar, and Xiaojin Zhu. 2017. Are key-foreign key joins safe to avoid when learning high-capacity classifiers? arXiv preprint arXiv:1704.00485 (2017)

    2017 arXiv

  28. [36]

    Carl Van Walraven, Irfan A Dhalla, Chaim Bell, Edward Etchells, Ian G Stiell, Kelly Zarnke, Peter C Austin, and Alan J Forster. 2010. Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. Cmaj 182,...

    2010

  29. [37]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  30. [38]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  31. [39]

    Jiayi Wang, Chengliang Chai, Nan Tang, Jiabin Liu, and Guoliang Li. 2022. Core- sets over multiple tables for feature-rich and data-efficient machine learning. Proceedings of the VLDB Endowment 16, 1 (2022), 64–76

    2022

  32. [40]

    Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei Wu, and Tat-Seng Chua

  33. [41]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 (2018)

    2018 arXiv

  34. [42]

    Lukáš Zahradník, Jan Neumann, and Gustav Šír. 2023. A deep learning blueprint for relational databases. In NeurIPS 2023 Second Table Representation Learning Workshop

    2023

  35. [43]

    Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep learning based recom- mender system: A survey and new perspectives. ACM computing surveys (CSUR) 52, 1 (2019), 1–38

    2019

  36. [44]

    Zixuan Zhao and Raul Castro Fernandez. 2022. Leva: Boosting machine learning performance with relational embedding data augmentation. In Proceedings of the 2022 International Conference on Management of Data . 1504–1517

    2022

  37. [2017]

    arXiv preprint arXiv:1708.04617 (2017)

    Attentional factorization machines: Learning the weight of feature interac- tions via attention networks. arXiv preprint arXiv:1708.04617 (2017)

    2017 arXiv

  38. [2022]

    In 2022 IEEE 38th International Conference on Data Engineering (ICDE)

    Feature augmentation with reinforcement learning. In 2022 IEEE 38th International Conference on Data Engineering (ICDE) . IEEE, 3360–3372

    2022

  39. [2024]

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

    AutoFeat: Transitive Feature Discovery over Join Paths. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 1861–1873

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.