Pith. sign in

REVIEW 3 major objections 6 minor 78 references

A column's joinability can be decided from 64 representative values, not all of them.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:42 UTC pith:FNALBBTX

load-bearing objection MosaicJoin is a real step forward for value-level semantic join discovery, but the core sketch-approximation claim is unproven and the non-metric cosine issue makes it load-bearing. the 3 major comments →

arxiv 2607.21781 v1 pith:FNALBBTX submitted 2026-07-23 cs.DB cs.AI

MosaicJoin: Compact Semantic Sketches for Value-Level Join Discovery

Grace Fan , Eden Wu , Majid Daliri , Juliana Freire This is my paper
classification cs.DB cs.AI
keywords semantic join discoveryvalue-level joinabilitydata lakessemantic sketchesk-center farthest-firstChamfer similarityquery subsamplingtop-k column retrieval
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MosaicJoin sets out to dissolve the trade-off in semantic join discovery between value-level accuracy and column-level speed. Its claim is that two columns can be judged joinable by comparing query values not against every value in a candidate column, but against a fixed-size sketch of representative values chosen by a k-center farthest-first procedure. The paper introduces a directed Chamfer-style score that averages, over query values, the best semantic match found in the sketch, so online cost is bounded by sketch size rather than column cardinality. A query-subsampling step with a Hoeffding-style error bound further cuts search time. If correct, this makes sub-second semantic join retrieval feasible over data lakes with up to 99,000 columns and a million values, without training or fine-tuning, and the paper reports up to 66x speedups over exact value-level comparison while maintaining or beating published methods on six benchmarks.

Core claim

MosaicJoin's claim is that a column's joinability can be faithfully summarized by a small set of actual value embeddings rather than by all values or by a single column vector. For each column, the method selects m representatives by farthest-first traversal—seeding at the embedding closest to the column's centroid, then repeatedly adding the embedding farthest from those already chosen—approximating the k-center objective of minimizing the worst-case distance from any value to its nearest representative. At query time it scores each candidate with a directed Chamfer-style score: the average over query values of the maximum cosine similarity against sketch embeddings. The paper argues this p

What carries the argument

The central object is the semantic sketch: a fixed-size subset S(C) of a column's value embeddings, built offline by greedy farthest-first k-center selection. It carries the argument because the maximum similarity over the sketch is treated as a proxy for the maximum over the full column, turning joinability scoring from O(n * |C|) into O(n * m). A second mechanism is the query-subsampling estimator, which scores only b randomly sampled query values and supplies the paper's formal error guarantee.

Load-bearing premise

The load-bearing premise is that 64 representative values chosen by k-center preserve each value's nearest-match similarity to the full column; if a joinable value falls far from every sketch point, its best match is underestimated and the column can drop out of the top-k.

What would settle it

Compute the exact directed Chamfer score over all candidate values and the sketch-based score with m=64 on a benchmark with deliberately rare joinable values, then compare top-10 rankings. If a candidate column whose only close matches are rare values is ranked lower by the sketch while exhaustive matching ranks it high, the coverage assumption fails. More directly, for any column, measure the fraction of values whose distance to their nearest sketch representative exceeds a threshold and test whether those values ever carry the join.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the sketch preserves Chamfer-style coverage, top-k semantic join retrieval runs at a cost fixed by sketch size, independent of candidate-column cardinality.
  • Query subsampling with b=1024 gives accuracy close to using the full query column while running roughly 4.5x faster, consistent with the paper's provable error bound.
  • MosaicJoin's reported gains over published methods—up to 17.6% in NDCG@20 and 135.7% in Precision@10—come without training, so new columns can be ingested by independently embedding and sketching them.
  • On the largest benchmark, the paper reports 0.32 seconds per query on columns with up to 57,000 query values and 1,000,000 lake values, versus 15.65 seconds for exact value-level search, making interactive and pipeline use plausible.
  • The method scales robustly as query columns grow: accuracy stays stable up to 57K query values, where published baselines degrade, because query subsampling bounds online cost regardless of query size.

Where Pith is reading between the lines

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

  • Editorial inference: the formal guarantee in the paper covers only the gap between the subsampled score and the full sketch score; the gap between sketch score and full-column score is assumed rather than proved. On long-tail columns, rare joinable values can fall outside the sketch, so the real accuracy loss from sketching is unmeasured.
  • Editorial inference: because the directed Chamfer score lets every query value take its best match, a candidate with a few very generic values can score high even when most of the column does not align. Precision-sensitive deployments may need a containment or uniqueness penalty that the paper does not explore.
  • Editorial inference: the paper's own ablation ties retrieval quality to embedding cluster compactness, suggesting the approach is a wrapper around embedding quality. Replacing the default low-dimensional text-embedding model with a stronger one should shift the accuracy-efficiency frontier, and this is testable without changing the sketch machinery.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. MosaicJoin proposes a training-free, value-level semantic join discovery method that represents each data-lake column as a fixed-size k-center sketch of value embeddings and scores candidate columns at query time by a Chamfer-style similarity computed against the sketch, with query values subsampled to a fixed budget. The paper claims this sketch-based approximation preserves full-column joinability while reducing online cost from 15.65s to 0.32s per query on high-cardinality benchmarks, and reports consistent improvements over DeepJoin, Snoopy, PEXESO, and other baselines on six benchmarks, including WDC-augmented variants.

Significance. If the central approximation claim were established, MosaicJoin would be a practically valuable contribution: it addresses a real trade-off between value-level accuracy and column-level efficiency, requires no task-specific training, and comes with a simple, correct Hoeffding-style bound for its query-subsampling estimator. The empirical study is broad, includes an exact-value baseline, and makes a credible attempt at scaling to 99K columns. However, the paper's load-bearing assertion—that a fixed-size k-center sketch approximates the full-column directed Chamfer score—is not proven, and the stated covering argument fails for cosine distance, which is not a metric. The evaluation also has a hyperparameter-selection leak on the headline benchmark and relies on LLM-generated silver labels with limited human validation. These issues are fixable but require substantive revision.

major comments (3)
  1. [§3.2 and §4.3 (Eq. (3), Theorem 2)] The central claim that a k-center sketch preserves the directed Chamfer score is unproven and the given justification is invalid for cosine distance. The paper states that if each original embedding is 'well covered' by a representative, then nearest-match similarities can be approximated using the sketch, but the covering argument requires a triangle inequality of the form D(q,s*) ≤ D(q,x*) + D(x*,s*). For D(x,y)=1−cos(x,y) this is false: on the unit circle, D(0°,120°)=1.5 > D(0°,60°)+D(60°,120°)=1.0. Thus the k-center covering radius does not translate into a similarity-approximation bound. Theorem 2 bounds only the query-subsampling error relative to the sketch score, not the sketch error relative to the full-column score. A concrete failure mode: a rare join-critical value absent from the m=64 sketch can yield near-zero sketch similarity for a query value that matches it exactly in t
  2. [§5.2 and §5.3 (Table 2, Table 3, Figure 3)] Hyperparameters—sketch size m=64, query sample size b=1024, threshold τ=0.1, embedding model, and the average-Chamfer variant—are selected by ablations on the Freyja benchmark and then the same benchmark is used to report the main state-of-the-art results. No held-out validation split, cross-validation, or multiple-seed analysis is used for the stochastic components (query subsampling, LLM labels). This evaluation leakage makes the reported improvements (e.g., NDCG@20 of 0.848 vs. DeepJoin's 0.777 on Freyja) difficult to interpret as robust out-of-sample gains. Please either tune on a validation subset distinct from the test set, use nested cross-validation, or demonstrate that the chosen settings are stable across all benchmarks.
  3. [§5.1.4 and Table 3] The expanded benchmark ground truth is generated by Gemini 2.5 Pro silver labels, with a validation sample of only 200 pairs. The audit checks agreement among three LLMs and three human curators on those 200 pairs, but the vast majority of labels are never directly validated. The power calculation about detecting 90%→95% precision does not address the correctness of the actual labels. The very low absolute Precision@10 on Webtables+WDC (MosaicJoin 0.066) suggests label noise or a very difficult search space. Since the expanded benchmarks are used to claim superiority over baselines (e.g., 135.7% relative improvement), the labeling methodology is load-bearing. Please provide a human-annotated gold sample for each expanded benchmark, or explicitly mark these results as pilot/exploratory.
minor comments (6)
  1. [Theorem 2 (title)] Typo: 'qery-subsampled' should be 'query-subsampled'.
  2. [§2, Related Work] The sentence 'we use DeepJoin as the representative column-level baseline' is duplicated verbatim in the same paragraph.
  3. [§2, Multivector Search] 'while admitting efficient top-k retrieval through through indexing' contains a duplicated 'through'.
  4. [Figure 7 caption] The caption says 'The bottom row subsamples values from the largest query bucket' but the figure appears to show a curve; please clarify what 'bottom row' refers to and how the sampling ratio is encoded.
  5. [§4.2] 'T o capture' has a stray space; also the symbol 'Φ(C)' is used both as a multiset and a set, which may confuse the discussion of duplicates in high-cardinality columns.
  6. [§2, Related Work] The name 'W arpGate' contains an erroneous space; should be 'WarpGate'.

Circularity Check

0 steps flagged

No significant circularity: the sketch approximation gap is a soundness issue, not a definitional or self-citation reduction.

full rationale

The paper's derivation chain is not circular. The semantic joinability score Jch (Eq. 2) is a defined measure; the sketch construction (Eq. 3) is a k-center selection rule; and Theorem 2 is a self-contained Hoeffding bound for subsampled estimation of the sketch score. None of these steps defines the claimed output in terms of itself or fits a parameter and then relabels it as a prediction. The load-bearing approximation claim in Section 3.2 ('if each original embedding is well covered by a nearby representative, then nearest-match similarities against the full column can be approximated using only the sketch') is asserted without a proof that relates sketch score to full-column score, but this is a correctness/soundness gap, not circularity; the conclusion even concedes that 'subsampling may overlook rare joinable values in extremely skewed distributions.' The only self-citations (e.g., the authors' technical report [24] for deferred indexing experiments on Webtables and Freyja) are not load-bearing for the main results, which are evaluated against published baselines and external benchmarks. Hyperparameters m=64 and b=1024 are selected from Freyja ablations and then used in Freyja results, which is an evaluation-leakage concern, but it does not make the central claim equal to a fitted constant. The sketch-to-full-column gap is real but belongs to correctness risk, not circularity, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The method introduces no new physical entities. It introduces a representation (semantic sketch) and an evaluation protocol, but no new particle, force, or conserved quantity. The main uncharged assumptions are embedding quality, sketch fidelity, and the Chamfer definition of joinability.

free parameters (4)
  • sketch size m = 64
    Selected via ablation on Freyja (Table 2); larger m gives marginal gains, smaller m is faster. Used in all main experiments, including the Freyja results.
  • query subsample size b = 1024
    Selected from the Freyja ablation (Figure 3); NDCG stabilizes around 1024 and runtime is improved. Used in all main experiments.
  • score threshold tau = 0.1
    Fixed in Section 4.3 as 'we use tau = 0.1' with no ablation reported; affects false positives and ranking.
  • embedding model = EmbeddingGemma 128-d
    Selected on Freyja as the best accuracy/efficiency tradeoff; BGE was more accurate but slower (Section 5.2).
axioms (6)
  • domain assumption Pretrained value embeddings place semantically equivalent surface forms close in R^d
    The whole semantic joinability definition replaces string equality with embedding similarity (Section 3.1). If the embedding model fails on abbreviations, aliases, or domain-specific vocabulary, recall collapses. The paper tests three embedding models but only on Freyja.
  • ad hoc to paper A fixed-size k-center sketch preserves the directed Chamfer score of the full column
    Section 3.2 asserts 'similarity computed against the sketch closely approximates similarity computed against all values' without a bound. No theorem relates J_ch(S(C), .) to J_ch(Phi(C), .). This is the load-bearing approximation.
  • domain assumption Directed many-to-one Chamfer aggregation is the correct operational definition of semantic joinability
    Section 3.1 defines joinability as average best-match similarity, justified by SEMA-JOIN many-to-one mappings. This excludes set-containment and one-to-one matching criteria used by SILKMOTH and KOIOS.
  • standard math Hoeffding's inequality applies to the subsampled estimator
    Theorem 2 proof; embeddings are normalized so each term lies in [-1,1], and indices are sampled independently. Mathematically valid.
  • standard math Farthest-first traversal is a constant-factor approximation to k-center in metric spaces
    Section 3.2 relies on this standard property, cited to [36] rather than the original source.
  • domain assumption LLM-as-a-judge silver labels and the 200-pair audit are accurate enough to serve as ground truth for expanded benchmarks
    Section 5.1.4 uses Gemini-generated labels with a multi-model audit on 200 pairs. No full label set is released, and the reported power calculation does not replace an actual significance test.

pith-pipeline@v1.3.0-alltime-deepseek · 22752 in / 14809 out tokens · 147562 ms · 2026-08-01T06:42:35.684446+00:00 · methodology

0 comments
read the original abstract

Join discovery is a core task in dataset search, enabling users to find columns that can be joined with a given query column. Early approaches focused on equi-joins, but data lakes and open-data repositories often contain columns whose values refer to the same entity but use different syntactic representations. To address this challenge, recent approaches discover semantically joinable columns but face a fundamental trade-off: methods that perform value-level comparisons accurately identify joinable columns but scale poorly to columns with high cardinality; column-level methods that encode an entire column into a single embedding are efficient but do not capture the fine-grained value alignment that determines whether a join is possible. We present MosaicJoin, a value-level semantic join discovery method that balances this trade-off. MosaicJoin achieves scalability through a novel sketching strategy that approximates the joinability of a column pair without having to compare all values. At query time, MosaicJoin scores each candidate sketch using a joinability score at a cost bounded by the sketch size, making retrieval efficient even for high-cardinality columns. A query subsampling operator further reduces online search time with provable accuracy guarantees, enabling robust retrieval for large query columns. Extensive experiments show that MosaicJoin outperforms previously published methods across all benchmarks while running up to 66 times faster than other value-level methods. MosaicJoin requires no training or fine-tuning, and it scales robustly to query columns containing up to 57K values and data lake columns containing up to 1M values.

Figures

Figures reproduced from arXiv: 2607.21781 by Eden Wu, Grace Fan, Juliana Freire, Majid Daliri.

Figure 1
Figure 1. Figure 1: Tradeoff between Accuracy of Value-level Joinabil￾ity and Online Search Efficiency of Column-level joinability, for query columns with up to 57K values (Section 5). reduces online search time from 15.65 seconds to 0.32 seconds while preserving high accuracy. This sub-second response time is impor￾tant both for interactive dataset search [7, 32], where users explore join candidates in real time, and for aut… view at source ↗
Figure 2
Figure 2. Figure 2: MosaicJoin Pipeline, with real columns from our experimental evaluations [48] (Section 5). During the offline phase, MosaicJoin embeds column values and creates a compact semantic sketch using the 𝑘-center procedure. During online processing, given a query column, MosaicJoin embeds the column values and ranks candidate columns using a Chamfer-style semantic joinability score. MosaicJoin correctly returns t… view at source ↗
Figure 3
Figure 3. Figure 3: Varying 𝑏 (number of sampled values from query columns) on Freyja benchmark We observe that NDCG@10 quickly stabilizes around 𝑏 = 1024, with little improvement beyond this point, while runtime is 4.5× faster than using the full query column. This indicates that random query subsampling preserves retrieval quality while improving effi￾ciency. Based on this tradeoff, we use 𝑏 = 1024 query samples in all expe… view at source ↗
Figure 4
Figure 4. Figure 4: UMAP projections of Freyja values embeddings from BGE, EmbeddingGemma, and MPNet Models. All em￾beddings are in gray. Red and blue points represent expected column values for two distinct queries. though such symmetry is not necessarily required in our inherently directed retrieval task. In all three cases, we use averages to keep scores comparable across query columns with different numbers of values and … view at source ↗
Figure 5
Figure 5. Figure 5: Effectiveness of MosaicJoin and baselines for vary￾ing 𝑘’s on Auto-FuzzyJoin, Webtables, and Freyja. but has significantly lower accuracy performance. We therefore select EmbeddingGemma as the default embedding model because it offers a much stronger efficiency–accuracy tradeoff. A key reason this works well is that EmbeddingGemma is trained with Matryoshka Representation Learning [44], which concentrates … view at source ↗
Figure 6
Figure 6. Figure 6: Effectiveness of MosaicJoin and baselines for vary￾ing 𝑘’s on the expanded benchmarks. thus producing false positives when a candidate column has a simi￾lar topic but is not value-level joinable. In contrast, Snoopy projects columns onto learned proxy columns to determine column join￾ability and to capture fine-grained value alignment. However, its proxies are trained on the WDC corpus, which may not captu… view at source ↗
Figure 7
Figure 7. Figure 7: Analysis of Precision@10 and NDCG@10 of Mo￾saicJoin, ESJ, DeepJoin, Snoopy, PEXESO as we vary the number of rows in query columns in the Freyja benchmark. Recall@10 follows a similar trend. The bottom row subsam￾ples values from the largest query bucket (up to 57K rows) in MosaicJoin; dashed lines show the performance of the baselines. Snoopy, and PEXESO gradually decrease. Among the baselines, PEXESO show… view at source ↗
Figure 8
Figure 8. Figure 8: Effectiveness of MosaicJoin with an index on Auto￾FuzzyJoin, compared to KOIOS, SILKMOTH, TOPJoin. much faster query search time. MosaicJoin constructs 𝑘-center se￾mantic sketches for each data lake column that maximize coverage of the column’s embedding space, producing a compact represen￾tation that captures the full semantics of even high-cardinality columns and enables Chamfer joinability scoring at a … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

78 extracted references · 7 linked inside Pith

  1. [1]

    Adelfio and Hanan Samet

    Marco D. Adelfio and Hanan Samet. 2013. Schema Extraction for Tabular Data on the Web. Proc. VLDB Endow. 6, 6 (2013), 421–432

    2013

  2. [2]

    Aline Bessa, Juliana Freire, Tamraparni Dasu, and Divesh Srivastava. 2020. Ef- fective discovery of meaningful outlier relationships. ACM Transactions on Data Science 1, 2 (2020), 1–33

    2020

  3. [3]

    Alex Bogatu, Alvaro A. A. Fernandes, Norman W. Paton, and Nikolaos Konstanti- nou. 2020. Dataset Discovery in Data Lakes. In ICDE. 709–720

    2020

  4. [4]

    Dan Brickley, Matthew Burgess, and Natasha F. Noy. 2019. Google Dataset Search: Building a search engine for datasets in an open Web ecosystem. In WWW. 1365–1375

    2019

  5. [5]

    Cafarella, Alon Y

    Michael J. Cafarella, Alon Y. Halevy, and Nodira Khoussainova. 2009. Data Integration for the Relational Web. Proc. VLDB Endow. 2, 1 (2009), 1090–1101

    2009

  6. [6]

    Riccardo Cappuzzo, Aimee Coelho, Félix Lefebvre, Paolo Papotti, and Gaël Varo- quaux. 2025. Retrieve, Merge, Predict: Augmenting Tables with Data Lakes. TMLR (2025)

    2025

  7. [7]

    Sonia Castelo, Rémi Rampin, Aécio S. R. Santos, Aline Bessa, Fernando Chirigati, and Juliana Freire. 2021. Auctus: A Dataset Search Engine for Data Discovery and Augmentation. Proc. VLDB Endow. 14, 12 (2021), 2791–2794

    2021

  8. [8]

    Adriane Chapman, Elena Simperl, Laura Koesten, George Konstantinidis, Luis- Daniel Ibáñez, Emilia Kacprzak, and Paul Groth. 2020. Dataset search: a survey. VLDB J. 29, 1 (2020), 251–272

    2020

  9. [9]

    Nadiia Chepurko, Ryan Marcus, Emanuel Zgraggen, Raul Castro Fernandez, Tim Kraska, and David R. Karger. 2020. ARDA: Automatic Relational Data Augmentation for Machine Learning. Proc. VLDB Endow. 13, 9 (2020), 1373–1387

    2020

  10. [10]

    Fernando Chirigati, Harish Doraiswamy, Theodoros Damoulas, and Juliana Freire. 2016. Data polygamy: the many-many relationships among urban spatio- temporal data sets. In SIGMOD. 1011–1025

    2016

  11. [11]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multi- modality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Tianji Cong, James Gale, Jason Frantz, H. V. Jagadish, and Çagatay Demiralp

  13. [13]

    Tianji Cong, Fatemeh Nargesian, and HV Jagadish. 2023. Pylon: Semantic Table Union Search in Data Lakes. arXiv preprint arXiv:2301.04901 (2023)

    Pith/arXiv arXiv 2023

  14. [14]

    Arash Dargahi Nobari and Davood Rafiei. 2024. DTT: An example-driven tabular transformer for joinability by leveraging large language models. Proc. ACM Manag. Data 2, 1 (2024), 1–24

    2024

  15. [15]

    Dong Deng, Albert Kim, Samuel Madden, and Michael Stonebraker. 2017. Silk- Moth: An Efficient Method for Finding Related Sets with Maximum Matching Constraints. Proc. VLDB Endow. 10, 10 (2017), 1082–1093

    2017

  16. [16]

    Yuhao Deng, Chengliang Chai, Lei Cao, Qin Yuan, Siyuan Chen, Yanrui Yu, Zhaoze Sun, Junyi Wang, Jiajun Li, Ziqi Cao, et al. 2024. Lakebench: A benchmark for discovering joinable and unionable tables in data lakes. Proc. VLDB Endow. 17, 8 (2024), 1925–1938

    2024

  17. [17]

    Yuyang Dong and Masafumi Oyamada. 2022. Table enrichment system for machine learning. In SIGIR. 3267–3271

    2022

  18. [18]

    Yuyang Dong, Kunihiro Takeoka, Chuan Xiao, and Masafumi Oyamada. 2021. Efficient joinable table discovery in data lakes: A high-dimensional similarity- based approach. In ICDE. IEEE, 456–467

    2021

  19. [19]

    Yuyang Dong, Chuan Xiao, Takuma Nozawa, Masafumi Enomoto, and Masafumi Oyamada. 2023. DeepJoin: Joinable Table Discovery with Pre-Trained Language Models. Proc. VLDB Endow. 16, 10 (2023), 2458–2470

    2023

  20. [20]

    Mahdi Esmailoghli, Jorge-Arnulfo Quiané-Ruiz, and Ziawasch Abedjan. 2022. MATE: Multi-Attribute Table Extraction. Proc. VLDB Endow. 15, 8 (2022), 1684– 1696

    2022

  21. [21]

    Grace Fan and Juliana Freire. 2025. Hierarchical table semantics for exploratory table discovery. In Proceedings of the Workshop on Human-In-the-Loop Data Analytics. 1–7

    2025

  22. [22]

    Grace Fan, Jin Wang, Yuliang Li, and Renée J. Miller. 2023. Table Discovery in Data Lakes: State-of-the-art and Future Directions. In SIGMOD Conference Companion. ACM, 69–75

    2023

  23. [23]

    Grace Fan, Jin Wang, Yuliang Li, Dan Zhang, and Renée J. Miller. 2023. Semantics- aware Dataset Discovery from Data Lakes with Contextualized Column-based Representation Learning. Proc. VLDB Endow. 16, 7 (2023), 1726–1739

    2023

  24. [24]

    Grace Fan, Eden Wu, Majid Daliri, and Juliana Freire. 2026. Technical Report on MosaicJoin: Compact Semantic Sketches for Value-Level Join Discovery. https://github.com/gracefan2020/MosaicJoin/blob/main/technical_report.pdf

    2026

  25. [25]

    Farid, Alexandra Roatis, Ihab F

    Mina H. Farid, Alexandra Roatis, Ihab F. Ilyas, Hella-Franziska Hoffmann, and Xu Chu. 2016. CLAMS: Bringing Quality to Data Lakes. In SIGMOD. 2089–2092

    2016

  26. [26]

    Raul Castro Fernandez, Ziawasch Abedjan, Famien Koko, Gina Yuan, Samuel Madden, and Michael Stonebraker. 2018. Aurum: A Data Discovery System. In ICDE. 1001–1012

    2018

  27. [27]

    Raul Castro Fernandez, Jisoo Min, Demitri Nava, and Samuel Madden. 2019. Lazo: A cardinality-based method for coupled estimation of jaccard similarity and containment. In ICDE. 1190–1201

    2019

  28. [28]

    Benjamin Feuer, Yurong Liu, Chinmay Hegde, and Juliana Freire. 2024. ArcheType: A Novel Framework for Open-Source Column Type Annotation using Large Language Models. Proc. VLDB Endow. 17, 9 (2024), 2279–2292

    2024

  29. [29]

    Juliana Freire, Grace Fan, Benjamin Feuer, Christos Koutras, Yurong Liu, Eduardo Pena, Aécio Santos, Cláudio T Silva, and Eden Wu. 2025. Large language models for data discovery and integration: Challenges and opportunities. IEEE Data Engineering Bulletin (2025)

    2025

  30. [30]

    Google. 2026. Gemini 3.5 Flash. https://ai .google.dev/gemini-api/docs/models/ gemini-3.5-flash

    2026

  31. [31]

    Yuxiang Guo, Yuren Mao, Zhonghao Hu, Lu Chen, and Yunjun Gao. 2025. Snoopy: Effective and Efficient Semantic Join Discovery via Proxy Columns. IEEE Trans. Knowl. Data Eng. 37, 5 (2025), 2971–2985

    2025

  32. [32]

    Halevy, Flip Korn, Natalya Fridman Noy, Christopher Olston, Neoklis Polyzotis, Sudip Roy, and Steven Euijong Whang

    Alon Y. Halevy, Flip Korn, Natalya Fridman Noy, Christopher Olston, Neoklis Polyzotis, Sudip Roy, and Steven Euijong Whang. 2016. Goods: Organizing Google’s Datasets. In SIGMOD. 795–806

    2016

  33. [33]

    Yeye He, Kris Ganjam, and Xu Chu. 2015. Sema-join: joining semantically-related tables using big table corpora. Proceedings of the VLDB Endowment 8, 12 (2015), 1358–1369

    2015

  34. [34]

    Xuming Hu, Chuan Lei, Xiao Qin, Asterios Katsifodimos, Christos Faloutsos, and Huzefa Rangwala. 2025. POLYJOIN: Semantic Multi-key Joinable Table Search in Data Lakes. In NAACL. 384–395

    2025

  35. [35]

    Xuming Hu, Shen Wang, Xiao Qin, Chuan Lei, Zhengyuan Shen, Christos Falout- sos, Asterios Katsifodimos, George Karypis, Lijie Wen, and Philip S. Yu. 2023. Automatic Table Union Search with Tabular Representation Learning. In ACL. 3786–3800

    2023

  36. [36]

    Rajesh Jayaram, Laxman Dhulipala, Majid Hadian, Jason D Lee, and Vahab Mir- rokni. 2024. MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encoding. NeurIPS 37 (2024), 101042–101073

    2024

  37. [37]

    Aamod Khatiwada, Grace Fan, Roee Shraga, Zixuan Chen, Wolfgang Gatter- bauer, Renée J Miller, and Mirek Riedewald. 2023. SANTOS: Relationship-based semantic table union search. SIGMOD 1, 1 (2023), 1–25

    2023

  38. [38]

    Aamod Khatiwada, Harsha Kokel, Ibrahim Abdelaziz, Subhajit Chaudhury, Ju- lian Dolby, Oktie Hassanzadeh, Zhenhan Huang, Tejaswini Pedapati, Horst Samulowitz, and Kavitha Srinivas. 2025. Tabsketchfm: Sketch-based tabular representation learning for data discovery over data lakes. In ICDE. 1523–1536

    2025

  39. [39]

    Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In SIGIR. 39–48

    2020

  40. [40]

    Harsha Kokel, Aamod Khatiwada, Tejaswini Pedapati, Haritha Ananthakrishnan, Oktie Hassanzadeh, Horst Samulowitz, and Kavitha Srinivas. 2025. Evaluating Joinable Column Discovery Approaches for Context-Aware Search.arXiv preprint arXiv:2510.24599 (2025)

    arXiv 2025

  41. [41]

    Harsha Kokel, Aamod Khatiwada, Tejaswini Pedapati, Haritha Ananthakr- ishnan, Oktie Hassanzadeh, Horst Samulowitz, and Kavitha Srinivas. 2025. TOPJoin: A Context-Aware Multi-Criteria Approach for Joinable Column Search. arXiv:2507.11505 [cs.DB]

    Pith/arXiv arXiv 2025

  42. [42]

    Christos Koutras and Juliana Freire. 2026. StraTyper: Automated Semantic Type Discovery and Multi-Type Annotation for Dataset Collections. arXiv preprint arXiv:2602.04004 (2026)

    arXiv 2026

  43. [43]

    Christos Koutras, Jiani Zhang, Xiao Qin, Chuan Lei, Vassilis N Ioannidis, Christos Faloutsos, George Karypis, and Asterios Katsifodimos. 2025. OmniMatch: Join- ability discovery in data products. Proc. VLDB Endow. 18, 11 (2025), 4588–4601

    2025

  44. [44]

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and Ali Farhadi. 2024. Matryoshka Representation Learning. arXiv:2205.13147 [cs.LG]

    Pith/arXiv arXiv 2024

  45. [45]

    Facebook AI Research Lab. 2015. fastText: Library for fast text representation and classification. (2015). https://fasttext .cc/

    2015

  46. [46]

    Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, Sören Auer, et al . 2015. Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia. Semantic web 6, 2 (2015), 167–195

    2015

  47. [47]

    Oliver Lehmberg, Dominique Ritze, Robert Meusel, and Christian Bizer. 2016. A Large Public Corpus of Web Tables containing Time and Context Metadata. In WWW. 75–76

    2016

  48. [48]

    Peng Li, Xiang Cheng, Xu Chu, Yeye He, and Surajit Chaudhuri. 2021. Auto- fuzzyjoin: Auto-program fuzzy similarity joins without labeled examples. In SIGMOD. 1064–1076

    2021

  49. [49]

    Shiyuan Liu, Jianwei Wang, Xuemin Lin, Lu Qin, Wenjie Zhang, and Ying Zhang

  50. [50]

    Marc Maynou, Sergi Nadal, Raquel Panadero, Javier Flores, Oscar Romero, and Anna Queralt. 2026. FREYJA: Efficient join discovery in data lakes. IEEE Trans. Knowl. Data Eng. 38 (2026), 1–12

    2026

  51. [51]

    Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. UMAP: Uniform Manifold Approximation and Projection.Journal of Open Source Software 3, 29 (2018), 861. https://doi .org/10.21105/joss.00861

    doi:10.21105/joss.00861 2018

  52. [52]

    Renée J. Miller. 2018. Open Data Integration. Proc. VLDB Endow. 11, 12 (2018), 2130–2139

    2018

  53. [53]

    Miller, Fatemeh Nargesian, Erkang Zhu, Christina Christodoulakis, Ken Q

    Renée J. Miller, Fatemeh Nargesian, Erkang Zhu, Christina Christodoulakis, Ken Q. Pu, and Periklis Andritsos. 2018. Making Open Data Transparent: Data Discovery on Open Data. IEEE Data Eng. Bull. 41, 2 (2018), 59–70

    2018

  54. [54]

    Pranay Mundra, Jianhao Zhang, Fatemeh Nargesian, and Nikolaus Augsten. 2023. Koios: Top-k semantic overlap set search. In ICDE. 1531–1543

    2023

  55. [55]

    Pu, Bahar Ghadiri Bashardoost, Erkang Zhu, and Renée J

    Fatemeh Nargesian, Ken Q. Pu, Bahar Ghadiri Bashardoost, Erkang Zhu, and Renée J. Miller. 2023. Data Lake Organization. IEEE Trans. Knowl. Data Eng. 35, 1 (2023), 237–250

    2023

  56. [56]

    Pu, Erkang Zhu, Bahar Ghadiri Bashardoost, and Renée J

    Fatemeh Nargesian, Ken Q. Pu, Erkang Zhu, Bahar Ghadiri Bashardoost, and Renée J. Miller. 2020. Organizing Data Lakes for Navigation. In SIGMOD. 1939– 1950

    2020

  57. [57]

    Miller, Ken Q

    Fatemeh Nargesian, Erkang Zhu, Renée J. Miller, Ken Q. Pu, and Patricia C. Arocena. 2019. Data Lake Management: Challenges and Opportunities. Proc. VLDB Endow. 12, 12 (2019), 1986–1989

    2019

  58. [58]

    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 (2018), 813–825

    2018

  59. [59]

    Arash Dargahi Nobari and Davood Rafiei. 2022. Efficiently transforming tables for joinability. In ICDE. 1649–1661

    2022

  60. [60]

    OpenAI. 2026. GPT-5.5. https://developers.openai.com/api/docs/models/gpt-5 .5

    2026

  61. [61]

    Paul Ouellette, Aidan Sciortino, Fatemeh Nargesian, Bahar Ghadiri Bashardoost, Erkang Zhu, Ken Pu, and Renée J. Miller. 2021. RONIN: Data Lake Exploration. Proc. VLDB Endow. 14, 12 (2021), 2863–2866

    2021

  62. [62]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Andreas Müller, Joel Nothman, Gilles Louppe, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vander- plas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Per- rot, and Édouard Duchesnay. 2018. Scikit-learn: Machine ...

    Pith/arXiv arXiv 2018

  63. [63]

    Ritze, O

    D. Ritze, O. Lehmberg, R. Meusel, C. Bizer, and S. Zope. 2015. WDC Web Table Cor- pus. http://webdatacommons .org/webtables/2015/downloadInstructions.html

    2015

  64. [64]

    Aécio Santos, Aline Bessa, Fernando Chirigati, Christopher Musco, and Juliana Freire. 2021. Correlation sketches for approximate join-correlation queries. In SIGMOD. 1531–1544

    2021

  65. [65]

    Aécio S. R. Santos, Aline Bessa, Christopher Musco, and Juliana Freire. 2022. A Sketch-based Index for Correlated Dataset Search. In ICDE. 2928–2941

    2022

  66. [66]

    Aécio S. R. Santos, Flip Korn, and Juliana Freire. 2024. Efficiently Estimating Mutual Information Between Attributes Across Tables. In ICDE. 193–206

    2024

  67. [67]

    Halevy, Hongrae Lee, Fei Wu, Reynold Xin, and Cong Yu

    Anish Das Sarma, Lujun Fang, Nitin Gupta, Alon Y. Halevy, Hongrae Lee, Fei Wu, Reynold Xin, and Cong Yu. 2012. Finding related tables. In SIGMOD. 817–828

    2012

  68. [68]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. Mpnet: Masked and permuted pre-training for language understanding. NeurIPS 33 (2020), 16857–16867

    2020

  69. [69]

    Henrique Schechter Vera, Sahil Dua, Biao Zhang, Daniel Salz, Ryan Mullins, Sindhu Raghuram Panyam, Sara Smoot, Iftekhar Naim, Joe Zou, Feiyang Chen, et al. 2025. Embeddinggemma: Powerful and lightweight text representations. arXiv preprint arXiv:2509.20354 (2025)

    Pith/arXiv arXiv 2025

  70. [70]

    WDC. [n.d.]. http://webdatacommons .org/webtables/goldstandard.html, last accessed on Feb 15, 2026

    2026

  71. [71]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]

    Pith/arXiv arXiv 2023

  72. [72]

    Haoxiang Zhang, Yurong Liu, Aécio Santos, Wei-Lun (Allen) Hung, and Juliana Freire. 2026. AutoDDG: Automated Dataset Description Generation using Large Language Models. Proc. ACM Manag. Data 4, 1, Article 12 (April 2026), 27 pages. https://doi.org/10.1145/3786626

    doi:10.1145/3786626 2026

  73. [73]

    Zixuan Zhao and Raul Castro Fernandez. 2022. Leva: Boosting Machine Learning Performance with Relational Embedding Data Augmentation. In SIGMOD. 1504– 1517

    2022

  74. [74]

    Erkang Zhu, Dong Deng, Fatemeh Nargesian, and Renée J. Miller. 2019. JOSIE: Overlap Set Similarity Search for Finding Joinable Tables in Data Lakes. In SIGMOD. 847–864

    2019

  75. [75]

    Erkang Zhu, Yeye He, and Surajit Chaudhuri. 2017. Auto-join: Joining tables by leveraging transformations. Proc. VLDB Endow. 10, 10 (2017), 1034–1045

    2017

  76. [76]

    Pu, and Renée J

    Erkang Zhu, Fatemeh Nargesian, Ken Q. Pu, and Renée J. Miller. 2016. LSH Ensemble: Internet-Scale Domain Search. Proc. VLDB Endow. 9, 12 (2016), 1185– 1196

    2016

  77. [2023]

    WarpGate: A Semantic Join Discovery System for Cloud Data Warehouses. In CIDR. 1–7

  78. [2026]

    arXiv preprint arXiv:2601.01015 (2026)

    HyperJoin: LLM-augmented Hypergraph Link Prediction for Joinable Table Discovery. arXiv preprint arXiv:2601.01015 (2026)

    arXiv 2026