Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Learning Relational Tabular Data without Shared Features

T0 review · 3 major / 5 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 The paper proposes treating training loss itself as the alignment signal for learning across tables with no shared features or pre-matched rows.

desk verdict Genuinely new problem and a promising empirical prototype, but the loss-alignment principle is not established — Theorem 4.1 is false on the reading that matters. read the letter →

arxiv 2502.10125 v1 pith:WQFPMZ37 submitted 2025-02-14 cs.LG cs.AI

Zhaomin Wu , Shida Wang , Ziyang Wang , Bingsheng He This is my paper · ORCID
classification cs.LGcs.AI MSC 68T0762J05
keywords tabulardatacross-tablelearningentityalignmentsoftattentionmechanismclustersamplingrelational
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 tries to establish that two relational tables can be learned together even when they share no feature and no row correspondence is given, by letting the drop in training loss identify which rows belong together. It formalizes this as the claim that properly aligned data achieve lower loss than misaligned data, proves the claim for linear regression in Theorem 4.1, and builds a system called Leal around a soft attention-based alignment coupled to a differentiable cluster sampler. If the claim holds for deep models, cross-table learning becomes possible in settings previously blocked by missing keys, with reported gains such as a 26.8% RMSE reduction on a house-price task.

What carries the argument

The load-bearing mechanism is loss-driven soft alignment implemented as multi-head attention: primary and secondary rows are mapped into a shared latent space, scaled inner products are softmax-normalized into alignment weights, and the weighted secondary embeddings are added to the primary representation before prediction. Coupled to it is a cluster sampler that keeps the candidate set small by assigning secondary rows to clusters with a Student's t soft k-means kernel and generating cluster weights from the primary row, then sampling $K$ candidates per primary row. Theorem 4.2 asserts this sampler can approximate any continuous optimal sampling function, which justifies replacing exact alignment search with learned sampling.

What would settle it

Take a synthetic pair of tables with known ground-truth alignment and enough model capacity to memorize: if training Leal on a randomly permuted secondary table reaches validation loss as low as, or lower than, training on the correctly aligned table, the loss-decay principle fails for deep learning and soft alignment weights carry no evidence of correspondence.

Watch

Extended reading notes

Core claim

The central discovery is that row alignment between tables can be learned from the optimization signal alone. For normalized feature matrices and a permutation matrix $R$, Theorem 4.1 states that the minimal squared error of the linear model $y = X_P \alpha + R X_S \beta$ under the correct alignment does not exceed the minimal squared error under a random alignment, so aligned data reach lower loss. Leal turns that observation into a mechanism: candidate rows from the secondary table are embedded, matched to a primary row by attention weights, and aggregated, with everything trained end-to-end by the supervised loss. On five real-world and five synthetic datasets, the paper reports that this soft alignment improves prediction over single-table baselines in most cases, most sharply on the house dataset where RMSE falls from 70.65 to 51.75.

Load-bearing premise

The whole framework rests on the unproven step that the loss advantage of aligned over misaligned rows, proven for linear regression, also holds for the deep networks Leal actually trains.

Editorial extensions

If this is right

  • If the loss-decay principle transfers to deep networks, supervised learning can ingest secondary tables that previously could not be joined at all because no key existed.
  • Soft alignment gives a self-supervised way to estimate match probabilities without labeled pairs, so alignment quality can improve as the prediction task is being solved.
  • The cluster sampler bounds the number of candidate pairs, making many-to-many fuzzy alignments feasible on large tables rather than requiring an explicit join.
  • The controlled experiments suggest that when the true matching row is among the candidates, the attention mechanism alone often recovers it, so improving candidate recall should directly improve the final model.

Reading between the lines

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

  • An implicit extension is that loss decay is a general correspondence signal: the same soft-alignment idea could be tried for other unpaired data forms, such as images and text, whenever a differentiable model can be trained end-to-end.
  • Because the theoretical guarantee covers only linear regression, a natural stress test is whether deep networks can memorize misaligned pairs and erase the loss gap; if they can, the attention weights are not evidence of true correspondence.
  • The reported bottleneck in clustering suggests the largest headroom for improvement is candidate recall, not the attention matcher itself, and a better trained clusterer might narrow the gap between Leal and perfect-alignment performance.
  • The training-loss gap could be reused as a cheap detector of whether two tables are related at all, which the paper identifies as future work for relationship discovery.
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

3 major / 5 minor

Summary. The paper introduces Leal, a framework for training on two relational tables with disjoint feature sets and no pre-aligned pairs. Leal scores candidate secondary records with an attention-based soft alignment, aggregates them into the primary representation, and uses a differentiable cluster sampler to limit the number of candidates. The central claim is that properly aligned data yield lower training loss than misaligned data; this is formalized for linear regression in Theorem 4.1, and the cluster sampler's expressive power is claimed in Theorem 4.2. Experiments compare Leal to single-table baselines on five real-world and five synthetic datasets, reporting up to a 26.8% RMSE reduction on the house dataset.

Significance. The problem is timely and the empirical setup is relevant. If valid, the loss-decay alignment principle would justify a genuinely new mechanism for cross-table learning without shared keys. The house dataset result (RMSE 51.75 vs. 70.65 for the best baseline) is a strong empirical finding, and the ablation in Figure 4, which checks attention capacity with a known ground-truth candidate, is a useful controlled experiment. However, the theoretical core of the paper is load-bearing and, in my assessment, incorrect; the contribution's significance therefore depends on whether the framework can stand without its central justification.

major comments (3)
  1. [Section 4.1, Appendix A.1 (Theorem 4.1)] Theorem 4.1 is not established and, on the reading required by Leal's use, false. The proof computes the OLS estimator for alpha in Eq. (10) using E[R] = (1/n)11^T in place of the realized permutation R, then cancels terms under the assumptions that X_S beta is orthogonal to the column space of X_P and that the residual is orthogonal to 11^T X_S beta; neither assumption is justified. More importantly, the statement fails for arbitrary finite-sample y: with n=4, X_P=[e1,e2], X_S=e3, and y=e3+2e4, where e_i are standard basis vectors in R^4 and all feature columns are unit-norm, the aligned OLS MSE is 1, while the average over uniform random permutations R of the misaligned MSE is 15/16 < 1. If the intended reading is instead that y is generated exactly as X_P alpha + R X_S beta for the correct R, then MSE_aligned=0 and the theorem is a tautology that does not cover the noisy finite-sample regime in which Leal is trained.
  2. [Section 4.1 (generalization to deep learning)] The justification for soft alignment is the loss-decay principle, and the paper explicitly states that extending Theorem 4.1 to deep learning remains an open challenge. Since Theorem 4.1 is invalid under the finite-sample reading, the empirical motivation in Figure 2 is the only support for the claim that deep networks reach lower loss on correctly aligned data. Figure 2 shows a simple linear-regression example and does not address memorization or shortcut solutions for the tabular settings used in the experiments. Without a valid theoretical or strong empirical link, the attention weights in Leal have no demonstrated connection to true entity correspondence; this is load-bearing because the method is trained solely by reducing prediction loss.
  3. [Section 4.2, Definition A.1, Appendix A.2 (Theorem 4.2)] Theorem 4.2 is close to tautological and its proof is a sketch. Definition A.1 defines the optimal cluster sampler as the minimizer of the prediction loss used to train Leal, and Theorem 4.2 proves that the implemented cluster sampler can approximate this minimizer. Since the same loss is minimized during training, the theorem only establishes representational flexibility, not that gradient-based training recovers the optimal sampler. The proof reduces the problem to approximating a scalar function h(p,s) by sums of univariate neural-network factors and invokes universal approximation and Taylor expansions, but it does not construct the required dimensions d and C, nor does it handle the actual n2-dimensional sampling-probability output of the cluster sampler.
minor comments (5)
  1. [Appendix A.1, Eq. (11) and Eq. (16)] The notation '1^T1' is dimensionally wrong for the expectation of a permutation matrix; the correct n x n matrix is 11^T.
  2. [References] The references for covertype, superconduct, letter, and breast contain the placeholder '[Insert date here]' and should be completed.
  3. [Section 5.1 and Section 5.3] There is a typo 'Environement' in Section 5.1, and Section 5.3 contains the ungrammatical phrase 'a the same table scale'.
  4. [Algorithm 1, line 8] Line 8 writes w_i but the loop index is b; it should be w_b to match Eq. (6).
  5. [Table 4 and Section 5.2] On the letter dataset, Leal (87.57%) is below Solo-ResNet (89.55%); the summary statement in Section 5.2 is accurate as 'majority' but could state this exception explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 4.1's 'aligned beats misaligned' proof reduces by construction to the trivial superset-of-parameters inequality.

  1. renaming known result [Appendix A.1, proof of Theorem 4.1, Eqs. (12)-(27); invoked in Section 4.1.]
    "The minimum is attained at β = 0, yielding: MSEmisaligned = ||(I − XP(XP⊤XP)−1XP⊤)y||2_2 ... Comparing with Equation 9, we conclude: MSEmisaligned ≥ inf_{α∈RmP,β∈RmS} ||y − XPα − XSβ||2_2 = MSEaligned"

    In the proof, the misaligned objective is reduced (Eqs. 17-24) to a problem whose optimum is β=0, so MSE_misaligned equals the residual of regressing y on XP alone. MSE_aligned is the infimum of the same residual while allowing any β (Eq. 9). The concluding inequality is therefore just the monotonicity of an infimum when the parameter set is enlarged: a model allowed to use XSβ cannot have higher training MSE than one forced to β=0. The permutation matrix R and the notion of alignment play no role in the final comparison; 'misaligned' has been made, by construction, equivalent to 'secondary table ignored'.

full rationale

The cluster-sampler approximation theorem (Theorem 4.2) is not circular: the 'optimal cluster sampler' is defined as a loss-minimizer, but the theorem proves a sup-norm approximation result using a Taylor-series/universal-approximation argument that does not rely on the loss-minimizing property, so the conclusion is not equivalent to the definition. The soft-alignment validation in Figure 4 is also not circular, because it evaluates attention on candidate sets containing a ground-truth correctly aligned record, and the training signal is the prediction loss, not an alignment label. The self-citations to Wu et al. 2022/2024 are contextual and not load-bearing. The main circular step is Theorem 4.1: as proved, the misaligned MSE is forced to the primary-only regression residual (β=0), and the aligned MSE is the infimum over a strictly larger parameter set, so the inequality is true by construction rather than by any property of alignment. The paper explicitly states that extending the result to deep learning 'remains an open challenge', which means the deep Leal framework rests on an empirical heuristic; that is a limitation and a correctness risk, not itself circularity. Because the paper's central theoretical justification for the loss-alignment principle reduces to a trivial superset-of-parameters inequality, while the empirical evaluations and the ground-truth-based ablation provide independent support, the overall circularity score is 6 rather than 0.

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

Leal introduces no new physical or conceptual entity; the latent space, attention weights, and clusters are trainable components of a neural architecture. The framework rests on an assumed near-functional dependency between tables, an unproved deep version of the loss-decay alignment principle, and a hidden orthogonality condition inside Theorem 4.1. The cluster sampler theorem relies on standard universal approximation without a concrete construction.

free parameters (4)
  • K (candidate count) = selected from {1,5,10,20,100} per dataset
    Tuned on validation; directly controls the size of the alignment search and the reported scalability.
  • C (cluster count) = selected from {1,5,10,20,100} per dataset
    Tuned on validation; interacts with K and dataset size, as shown in ablation Figures 5-6.
  • Embedding dimension d = 100
    Set uniformly across experiments; not tuned per dataset.
  • Student t kernel degree of freedom gamma = not reported
    In Equation 5, the degree of freedom gamma is not specified or tuned, yet it shapes the in-cluster probabilities.
assumptions (4)
  • domain assumption Strong functional dependency IF(Y;X_S|X_P) is close to 1.
    The problem formulation in Section 3 assumes the secondary table carries nearly all residual information about the label; without this, no cross-table benefit is expected.
  • ad hoc to paper Loss minimization identifies correct alignments in deep networks.
    Section 4.1 explicitly leaves the deep extension open; the entire soft alignment mechanism depends on this unproved generalization.
  • ad hoc to paper X_S beta is orthogonal to the column space of X_P in the proof of Theorem 4.1.
    Appendix A.1 asserts without proof that the projection term vanishes; this is generally false and invalidates the derivation as written.
  • standard math Neural networks can approximate arbitrary continuous functions.
    Theorem 4.2 proof relies on universal approximation results, which are standard but not instantiated with explicit widths or constructions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Relational Tabular Data without Shared Features." pith.science (2026). https://pith.science/paper/WQFPMZ37

@misc{pith2026250210125,
  author       = {Pith},
  title        = {Pith review of: Learning Relational Tabular Data without Shared Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WQFPMZ37}},
  note         = {Machine review of arXiv:2502.10125}
}
read the original abstract

Learning relational tabular data has gained significant attention recently, but most studies focus on single tables, overlooking the potential of cross-table learning. Cross-table learning, especially in scenarios where tables lack shared features and pre-aligned data, offers vast opportunities but also introduces substantial challenges. The alignment space is immense, and determining accurate alignments between tables is highly complex. We propose Latent Entity Alignment Learning (Leal), a novel framework enabling effective cross-table training without requiring shared features or pre-aligned data. Leal operates on the principle that properly aligned data yield lower loss than misaligned data, a concept embodied in its soft alignment mechanism. This mechanism is coupled with a differentiable cluster sampler module, ensuring efficient scaling to large relational tables. Furthermore, we provide a theoretical proof of the cluster sampler's approximation capacity. Extensive experiments on five real-world and five synthetic datasets show that Leal achieves up to a 26.8% improvement in predictive performance compared to state-of-the-art methods, demonstrating its effectiveness and scalability.

Figures

Figures reproduced from arXiv: 2502.10125 by the authors.

Figure 1
Figure 1. Example of latent alignment learning in a financial appli￾cation: (left) bank transaction table and (right) Bitcoin transaction table. The tables lack shared features, and the label “Bitcoin￾related” can be inferred if bank withdrawals are proportional to Bitcoin deposits (or vice versa). et al., 2021; 2025; Chen & Guestrin, 2016; Prokhorenkova et al., 2018) primarily focus on learning from individual tables in isol… view at source ↗
Figure 2
Figure 2. Relationship between data alignment and model training loss. Consider the training of a model on two relational tables: the first table contains a single feature x1 and a binary label y, while the second table contains a single feature x2. The left subfigure illustrates the distribution of data points (x1, x2) when x1 and x2 are properly aligned, while the right subfigure depicts the case of misaligned data. Both su… view at source ↗
Figure 3
Figure 3. Overall model structure of Leal features, we first map them into a shared latent space: z P i = f P (x P i ), z S = f S (x S ) (1) In this latent space, the distance between z P i ∈ R 1×d and each row vector of z S ∈ R K×d represents the relation￾ship between x P i and each record in x S, where d is the dimensionality of the latent space, and K is the number of candidate records from the secondary table. The scaled … view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Effect of cluster size C on performance 10 0 10 1 10 2 Number of Samples (K) 200 300 400 500 RMSE Bike C=1 C=10 C=100 10 0 10 1 10 2 Number of Samples (K) 70 72 74 76 78 80 82 Accuracy (%) Hepatitis C=1 C=10 C=100 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 28 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Airbnb prices dataset, 2019

    Airbnb . Airbnb prices dataset, 2019. URL http://insideairbnb.com/get-the-data.html

    work page 2019

  3. [3]

    Ann-benchmarks: A benchmarking tool for approximate nearest neighbor algorithms

    Aum \"u ller, M., Bernhardsson, E., and Faithfull, A. Ann-benchmarks: A benchmarking tool for approximate nearest neighbor algorithms. Information Systems, 87: 0 101374, 2020

    work page 2020

  4. [4]

    Ba, J. L. Layer normalization. arXiv preprint arXiv:1607.06450, 2016

    arXiv 2016

  5. [5]

    Ecml/pkdd 2002 discovery challenge: Hepatitis data set, 2002

    Berka, P. Ecml/pkdd 2002 discovery challenge: Hepatitis data set, 2002. URL https://www.semanticscholar.org/paper/094b3baf58e58ca5aa89bf4913536139991d3b90

    work page 2002

  6. [6]

    Covertype, 1998

    Blackard, J. Covertype, 1998. URL https://doi.org/10.24432/C50K5N. Accessed: [Insert date here]

    doi:10.24432/c50k5n 1998

  7. [7]

    and Guestrin, C

    Chen, T. and Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pp.\ 785--794, 2016

    2016

  8. [8]

    Citi bike system data, 2016

    CitiBike . Citi bike system data, 2016. URL https://www.citibikenyc.com/system-data

    work page 2016

Show all 44 references
  1. [9]

    The faiss library

    Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazar \'e , P.-E., Lomeli, M., Hosseini, L., and J \'e gou, H. The faiss library. arXiv preprint arXiv:2401.08281, 2024

    2024 arXiv

  2. [10]

    Revisiting deep learning models for tabular data

    Gorishniy, Y., Rubachev, I., Khrulkov, V., and Babenko, A. Revisiting deep learning models for tabular data. Advances in Neural Information Processing Systems, 34: 0 18932--18943, 2021

    2021

  3. [11]

    Tabm: Advancing tabular deep learning with parameter-efficient ensembling

    Gorishniy, Y., Kotelnikov, A., and Babenko, A. Tabm: Advancing tabular deep learning with parameter-efficient ensembling. The Thirteenth International Conference on Learning Representations (ICLR), 2025

    2025

  4. [12]

    Gisette dataset, 2004

    Guyon, I., Gunn, S., Ben-Hur, A., and Dror, G. Gisette dataset, 2004. URL https://www.csie.ntu.edu.tw/ cjlin/libsvmtools/datasets/binary/gisette_scale.bz2

    2004

  5. [13]

    Superconductivity data, 2018

    Hamidieh, K. Superconductivity data, 2018. URL https://doi.org/10.24432/C53P47. Accessed: [Insert date here]

    2018 doi

  6. [14]

    Fedcvt: Semi-supervised vertical federated learning with cross-view training

    Kang, Y., Liu, Y., and Liang, X. Fedcvt: Semi-supervised vertical federated learning with cross-view training. ACM Transactions on Intelligent Systems and Technology (TIST), 13 0 (4): 0 1--16, 2022

    2022

  7. [15]

    H., Yatskar, M., Yin, D., Hsieh, C.-J., and Chang, K.-W

    Li, L. H., Yatskar, M., Yin, D., Hsieh, C.-J., and Chang, K.-W. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019

    1908 arXiv

  8. [16]

    H., You, H., Wang, Z., Zareian, A., Chang, S.-F., and Chang, K.-W

    Li, L. H., You, H., Wang, Z., Zareian, A., Chang, S.-F., and Chang, K.-W. Unsupervised vision-and-language pre-training without parallel images and captions. arXiv preprint arXiv:2010.12831, 2020

    2010 arXiv

  9. [17]

    Learnable fourier features for multi-dimensional spatial positional encoding

    Li, Y., Si, S., Li, G., Hsieh, C.-J., and Bengio, S. Learnable fourier features for multi-dimensional spatial positional encoding. Advances in Neural Information Processing Systems, 34: 0 15816--15829, 2021

    2021

  10. [18]

    Vertical federated learning: Concepts, advances, and challenges

    Liu, Y., Kang, Y., Zou, T., Pu, Y., He, Y., Ye, X., Ouyang, Y., Zhang, Y.-Q., and Yang, Q. Vertical federated learning: Concepts, advances, and challenges. IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  11. [19]

    Decoupled weight decay regularization

    Loshchilov, I. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  12. [20]

    Discovering reliable approximate functional dependencies

    Mandros, P., Boley, M., and Vreeken, J. Discovering reliable approximate functional dependencies. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp.\ 355--363, 2017

    2017

  13. [21]

    When do neural nets outperform boosted trees on tabular data? Advances in Neural Information Processing Systems, 36, 2024

    McElfresh, D., Khandagale, S., Valverde, J., Prasad C, V., Ramakrishnan, G., Goldblum, M., and White, C. When do neural nets outperform boosted trees on tabular data? Advances in Neural Information Processing Systems, 36, 2024

    2024

  14. [22]

    and Eich, M

    Mishra, P. and Eich, M. H. Join processing in relational databases. ACM Computing Surveys (CSUR), 24 0 (1): 0 63--113, 1992

    1992

  15. [23]

    J., Pu, K

    Nargesian, F., Zhu, E., Miller, R. J., Pu, K. Q., and Arocena, P. C. Data lake management: challenges and opportunities. Proceedings of the VLDB Endowment, 12 0 (12): 0 1986--1989, 2019

    1986

  16. [24]

    Tlc trip record data, 2016

    New York TLC . Tlc trip record data, 2016. URL https://www1.nyc.gov/site/tlc/about/tlc-trip-record-data.page

    2016

  17. [25]

    Bridging data silos using big data integration

    Patel, J. Bridging data silos using big data integration. International Journal of Database Management Systems, 11 0 (3): 0 01--06, 2019

    2019

  18. [26]

    V., and Gulin, A

    Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V., and Gulin, A. Catboost: unbiased boosting with categorical features. Advances in neural information processing systems, 31, 2018

    2018

  19. [27]

    Kaggle dataset: Housing price in beijing, 2017

    Qiu, Q. Kaggle dataset: Housing price in beijing, 2017. URL https://www.kaggle.com/ruiqurm/lianjia

    2017

  20. [28]

    and Nicolae, D

    Reimherr, M. and Nicolae, D. L. On quantifying dependence: A framework for developing interpretable measures. 2013

    2013

  21. [29]

    Secondary school rankings in singapore, 2020

    Salary.sg. Secondary school rankings in singapore, 2020. URL https://www.salary.sg/2020/secondary-schools-ranking-2020-psle-cut-off/

    2020

  22. [30]

    Resale flat prices in singapore, 2018

    Singapore HDB . Resale flat prices in singapore, 2018. URL https://data.gov.sg/dataset/resale-flat-prices

    2018

  23. [31]

    Letter recognition, 1991

    Slate, D. Letter recognition, 1991. URL https://doi.org/10.24432/C5ZP40. Accessed: [Insert date here]

    1991 doi

  24. [32]

    Prometna varnost (traffic safety) statistics

    Slovenian Police . Prometna varnost (traffic safety) statistics. URL https://web.archive.org/web/20170918151410/http://www.policija.si/index.php/statistika/prometna-varnost#expand

  25. [33]

    Vl-bert: Pre-training of generic visual-linguistic representations

    Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., and Dai, J. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019

    1908 arXiv

  26. [34]

    Sun, J., Xu, Z., Yang, D., Nath, V., Li, W., Zhao, C., Xu, D., Chen, Y., and Roth, H. R. Communication-efficient vertical federated learning with limited overlapping samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 5203--5212, 2023

    2023

  27. [35]

    Attention is all you need

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

    2017

  28. [36]

    Wikidbs: A large-scale corpus of relational databases from wikidata

    Vogel, L., Bodensohn, J.-M., and Binnig, C. Wikidbs: A large-scale corpus of relational databases from wikidata. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024

    2024

  29. [37]

    Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix

    Wang, T., Jiang, W., Lu, Z., Zheng, F., Cheng, R., Yin, C., and Luo, P. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pp.\ 22680--22690. PMLR, 2022

    2022

  30. [38]

    A coupled design of exploiting record similarity for practical vertical federated learning

    Wu, Z., Li, Q., and He, B. A coupled design of exploiting record similarity for practical vertical federated learning. Advances in Neural Information Processing Systems, 35: 0 21087--21100, 2022

    2022

  31. [39]

    Federated transformer: Multi-party vertical federated learning on practical fuzzily linked data

    Wu, Z., Hou, J., Diao, Y., and He, B. Federated transformer: Multi-party vertical federated learning on practical fuzzily linked data. In Advances in Neural Information Processing Systems, 2024

    2024

  32. [40]

    Unsupervised deep embedding for clustering analysis

    Xie, J., Girshick, R., and Farhadi, A. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pp.\ 478--487. PMLR, 2016

    2016

  33. [41]

    Xu, P., Zhu, X., and Clifton, D. A. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (10): 0 12113--12132, 2023

    2023

  34. [42]

    Ptarl: Prototype-based tabular representation learning via space calibration

    Ye, H., Fan, W., Song, X., Zheng, S., Zhao, H., Guo, D., and Chang, Y. Ptarl: Prototype-based tabular representation learning via space calibration. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR), 2024. URL https://openreview.net/foru...

    2024

  35. [43]

    Autoencoder and its various variants

    Zhai, J., Zhang, S., Chen, J., and He, Q. Autoencoder and its various variants. In 2018 IEEE international conference on systems, man, and cybernetics (SMC), pp.\ 415--419. IEEE, 2018

    2018

  36. [44]

    and Soklic, M

    Zwitter, M. and Soklic, M. Breast cancer, 1988. URL https://doi.org/10.24432/C51P4M. Accessed: [Insert date here]

    1988 doi

Pith tools

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