Pith. sign in

REVIEW 3 major objections 2 minor 34 references

The simplest sampling strategy — random uniform — beats all nine benchmarked progressive samplers for data-quality profiling on real datasets, including an 11–49x accuracy gap over dependency-guided MCMC.

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 02:38 UTC pith:S4IJD2TY

load-bearing objection Solid, reproducible empirical study showing IQR-proxy-guided samplers underperform random uniform, but the paper overreaches when it says 'representativeness, not domain knowledge'—the evidence only covers one proxy family. the 3 major comments →

arxiv 2607.25356 v1 pith:S4IJD2TY submitted 2026-07-28 cs.DB cs.AIcs.CL

Data Quality Profiling at Scale with Progressive Sampling: A Benchmark for Data-Centric AI Pipelines

Laure Berti-Equille This is my paper
classification cs.DB cs.AIcs.CL
keywords data quality profilingprogressive samplingrandom uniform samplingMCMC samplingdata-centric AIIQR proxytabular databenchmark
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.

Progressive sampling is how large tabular datasets are monitored for data quality without a full scan; the field's open question is which sampler preserves profile fidelity best. The paper compares nine strategies — blind draws (random uniform, geometric, Yamane, cluster) against proxy-guided ones (Metropolis-Hastings, DAG-guided MCMC, stratified and importance-weighted variants) — across synthetic data, three real administrative/census datasets, an IoT sensor stream, and ultra-large tables up to 7.4M rows. Its central claim is a negative one: blind representative samplers dominate uniformly. Random uniform reaches 0.49% mean relative error on the NYC 311 requests dataset at a 5% budget, while DAG-guided MCMC stays near 19–27% error regardless of budget and is 11–49x worse across real datasets. The stated reason is an IQR proxy mismatch: guided samplers target numeric outliers, but quality defects concentrate in categorical columns the proxy cannot see; the practical conclusion the author draws is that representativeness, not domain knowledge, decides sampler quality.

Core claim

The paper's core discovery is that, contrary to the intuition that exploiting attribute dependency structure should improve profile estimation, a schema-free random uniform sampler is the most accurate and scalable of the nine strategies for multi-indicator data quality profiling. On all real datasets, from 500K-row administrative tables to a 7.4M-row running-log table to a 2.3M-row sensor stream, random uniform and cluster sampling matched or beat every proxy-guided method, with DAG-guided MCMC 11–49x less accurate (Wilcoxon W=0, p=0.002, n=9 pairs) and 12x slower at 5M rows. The mechanism identified is the per-row IQR error proxy shared by all guided samplers: it over-selects numeric outli

What carries the argument

The load-bearing object is the per-row IQR error proxy σ_i used by all guided samplers, computed as the fraction of numeric columns in which the row is a Tukey outlier plus its missing-value fraction. This proxy feeds a Metropolis–Hastings acceptance rule and Horvitz–Thompson importance weights in the DAG-guided, MH, stratified, and importance-weighted samplers, steering draws toward rows with high proxy scores. The paper's argument turns on showing that this proxy is structurally mismatched to real quality profiles: it cannot see categorical or string-typed columns, and on numeric-heavy IoT data it inflates the empirical IQR, collapsing the estimated outlier rate. The benchmark's comparison

Load-bearing premise

The negative result depends on the assumption that guided samplers will keep using a numeric-outlier proxy that ignores categorical columns; the author notes that a categorical-aware proxy is an open future avenue, so a better-guided sampler might change the ranking.

What would settle it

A concrete way to test the central claim: implement a guided sampler whose proxy is trained to flag categorical/string defects (missing values, functional-dependency violations, and malformed codes) rather than numeric outliers, and benchmark it against random uniform on the same administrative datasets at a 5% budget. If its mean relative error falls below random uniform's 0.49% on NYC 311 (or below the 0.004–0.020 range on other real datasets), the uniform-dominance claim is false. A weaker falsifier: rerun the correlation ablation with at least 20 seeds per correlation level; if DAG's direc

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the result holds, production data-quality gates in data-centric AI pipelines can run on random uniform or cluster sampling at a 5–10% budget and stay below 1% mean relative error on typical administrative/census data, with near-linear scaling.
  • Dependency-graph construction and per-row proxy scoring can be eliminated from profiling pipelines without accuracy loss, removing a schema-maintenance cost that grows as data evolves.
  • DAG's error remains near 19–27% regardless of budget, implying that increasing the sampling budget for a guided sampler does not fix profiling bias; the sampling mechanism must change instead of the budget.
  • On numeric-dominated IoT streams, proxy-guided samplers can report outlier rates near zero against a true rate of about 30%, so monitoring pipelines relying on them would miss genuine fault signals.
  • Random uniform is also the most robust baseline under injected missing, duplicate, outlier, and functional-dependency errors up to 30% injection rates, making it a safe default even in corrupted conditions.

Where Pith is reading between the lines

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

  • If a guided sampler used a proxy that detects categorical quality defects (e.g., missingness and FD violations in string columns), the ranking might change on administrative data; the paper itself leaves this as an open avenue, so testing such a proxy on NYC 311 would be a direct check on the generality of the negative result.
  • Cluster sampling's near-identical performance to random uniform suggests storage-order block structure is not a significant confounder in these datasets; this could become an advantage or liability depending on physical data ordering in other workloads.
  • The IQR estimation-collapse failure mode implies a testable design rule: any sample used to estimate IQR-based outlier rates must preserve the population quartiles; comparing sample Q1/Q3 against population values at growing budgets would verify this directly.
  • A cheap diagnostic for practitioners: before deploying any guided sampler, compare its profile estimate against random uniform on a held-out real dataset at the target budget; if the guided sampler is not clearly better, use random uniform.

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 / 2 minor

Summary. The paper benchmarks nine progressive sampling strategies for multi-indicator data quality profiling (missing rate, duplicate rate, outlier rate, FD-inconsistency rate) on real and synthetic datasets up to 7.4M rows, using exhaustive ground truth as the reference. The central claim is that blind representative samplers (random uniform, cluster) uniformly dominate proxy-guided samplers (DAG-guided MCMC, Metropolis-Hastings, stratified, importance-weighted) on real data, with DAG 11–49x worse in MRE; the authors attribute the failure to an IQR-based numeric-outlier proxy that is blind to categorical quality defects and conclude that 'representativeness, not domain knowledge, determines sampler quality.' The paper also reports scalability exponents, an IoT failure mode ('IQR estimation collapse'), and robustness under error injection. It provides a reproducibility package with code and archived data, and includes a controlled ablation (E7a) that isolates the weighting step.

Significance. If the central claim is accepted, this is a useful reproducible negative result: a specific family of structure-aware, IQR-proxy-guided row samplers adds computational overhead without accuracy benefit for multi-indicator quality profiling of administrative, census, and IoT tabular data. The exhaustive ground truth, public code/archived data, and the E7a ablation are notable strengths, as is the use of a Wilcoxon signed-rank test on independent dataset-seed pairs. However, the paper's headline conclusion is considerably broader than the evidence: every guided method uses the same IQR-based numeric proxy (Eq. 2), so the benchmark tests one proxy family, not 'domain knowledge' in general. The paper's own future-work note that a categorical-aware proxy is an open avenue, together with the dag_uniform ablation, limits the generality of the negative result.

major comments (3)
  1. [Section 4, Eq. (2), Table 14] The abstract and conclusion claim that 'representativeness, not domain knowledge, determines sampler quality.' The evidence supports only that IQR-proxy-guided samplers underperform blind uniform sampling on the tested datasets. All guided methods (MH, DAG, Gibbs, strat-col, strat-quality, importance) use the same Tukey-IQR numeric-outlier/missing-fraction proxy of Eq. (2); no categorical-quality-aware proxy is tested. Table 14 (E7a) shows that dag_uniform, which keeps DAG's graph proposal but drops IQR weights, beats RU on D3 at b=5% (0.024 vs 0.037) and cuts D2 b=50 MRE from 19.5% to 3.8%, locating the failure in the IQR weighting step, not in dependency structure or domain knowledge broadly. The future-work paragraph concedes that a proxy capturing categorical quality defects remains open. The strong claim should be rephrased to 'IQR-based proxy guidance' or supported by experiments w
  2. [Section 3.8, Table 13] The text states that on D7-real dag 'reports 26.9% MRE across all budgets including 100%'. At budget b=1, Definition 2 gives n=N, so the sample is the full dataset and any estimator should reproduce the exhaustive ground truth. A 26.9% MRE at b=1 would imply the algorithm does not actually return all rows, i.e., a bug in the progressive loop or profile computation. Table 13 only shows b=5% and b=50%, so this is likely a wording error, but as written it undercuts the 'IQR estimation collapse' mechanism, which cannot persist when the sample is the population. Please correct or clarify.
  3. [Section 3.1, Tables 5, 8, 12] Most experiments use 3 seeds and no variance or confidence intervals are reported for the headline MRE values. The RU-vs-DAG comparison is backed by a Wilcoxon test on n=9 pairs, but the '11–49x' and '28–47x' ratios are ratios of point estimates and can be unstable given the tiny RU denominators (e.g., D2 RU MRE 0.004 vs 0.003 changes the ratio from 49x to 65x). Report per-seed values or bootstrap CIs for the primary tables, and state explicitly that E5's N=5M point rests on a single seed.
minor comments (2)
  1. [Figures 1 and 2 captions] The captions mislabel dataset IDs: 'D4 (NYC 311)', 'D5 (NYPD)', and 'D7 (Adult)' should be D2, D3, and D4 respectively. This is confusing because Table 3 defines D2=NYC 311, D3=NYPD, D4=Adult, and D5 is synthetic scalability data.
  2. [Section 3.7, Table 12] The text says DAG requires 28.13s on D6-B at 5% budget, but Table 12 reports 27.58s. Please reconcile.

Circularity Check

0 steps flagged

No circular dependency: benchmark compares samplers against exhaustive ground truth; the IQR-proxy/DAG design is a stated limitation, not an input that forces the conclusion.

full rationale

The paper's central claim—that blind representative samplers outperform proxy-guided MCMC samplers—is obtained by comparing each sampler's estimated profile against the exhaustive ground-truth profile Q(D) (Definition 3, Eq. rel_err). No parameter is fitted to the target quality indicators, and the power-law exponents in E5 are descriptive fits, not predictions used to establish the main finding. The only self-referential element is that the author's own DAG sampler is used as the representative guided method, and the IQR proxy in Eq. 2 uses the same Tukey fence as the outlier-rate indicator in Definition 1. The paper explicitly states this alignment ('ensuring the proxy targets the same anomaly type it is asked to estimate'), but this does not make the result circular: aligning the proxy with the metric could only make the guided sampler more competitive on the outlier dimension, and the ground-truth comparison shows it is far worse. The paper also openly scopes the negative result to IQR-style proxies: the E7a ablation (dag_uniform 3.8% vs DAG 19.5% at b=50%) attributes the failure to the IQR weighting, and the Future Work section concedes that 'Stratified sampling with a proxy that captures categorical quality defects (not IQR-based) remains an open avenue.' These are honest boundary conditions, not circularity. Self-citations to related work (Abedjan et al., Berti-Équille et al.) are contextual and not load-bearing; no uniqueness theorem is imported, and the DAG sampler is explicitly described as an original design. The benchmark is self-contained against external ground truth, so no circular step is exhibited.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The benchmark's conclusion depends primarily on the choice of MRE metric and the shared IQR proxy; the hyperparameters are documented and partially sensitivity-checked.

free parameters (9)
  • MRE denominator floor = 0.05
    Chosen to match a 5% practical significance threshold; affects measured errors, especially on D1 where ground-truth rates are below floor.
  • MRE cap = 1.0
    Caps per-indicator relative error at 100%; prevents extreme outliers from dominating, but masks magnitude.
  • IQR fence multiplier = 1.5
    Used both in the outlier indicator and the sampling proxy; the paper notes a 3xIQR fence would alter DAG-MH ordering on D4, so the negative result is sensitive to this choice.
  • DAG correlation threshold rho_min = 0.3
    Edges added if |Pearson r|>0.3; no sensitivity analysis reported.
  • MCMC temperature lambda = 2.0
    Metropolis acceptance temperature; no sensitivity reported.
  • MCMC batch size B = 500 (E1-E4, E8) / 1000 (E3, E5)
    Swept for DAG on D2 (B in {100,200,500,1000}) with MRE variation <0.001, so low risk.
  • Progressive growth factor gamma = 2
    Standard geometric schedule from Provost et al.; not tuned.
  • Convergence epsilon and patience k = 0.01, 1
    Algorithm 2 stopping criteria; benchmark uses b0=0.05 and these defaults.
  • Cluster block count k = max(10, floor(sqrt(N)))
    Square-root rule for number of contiguous blocks; not tuned.
axioms (5)
  • domain assumption Exhaustive ground-truth profiles are correct for all datasets
    The benchmark relies on full-scan profiles as ground truth; any errors in those profiles would propagate.
  • domain assumption The four quality indicators (missing, duplicate, outlier, FD violation) constitute a sufficient profile for DC-AI pipelines
    If the profile matters differently (e.g., other defect types), the ranking of samplers could change.
  • domain assumption The MRE metric with floor/cap is a fair summary of profiling accuracy
    The floor at 0.05 makes near-zero rates contribute only absolute deviations; this shapes the D1 results and cross-dataset comparisons.
  • ad hoc to paper The IQR-based error proxy is a representative implementation of proxy-guided sampling
    The DAG, MH, stratified, and importance samplers all share this proxy; the paper's negative result is a property of this proxy family, as it acknowledges that a categorical-aware proxy is an open avenue.
  • domain assumption Datasets are representative of production tabular data in DC-AI settings
    Six real datasets from administrative, census, IoT domains support the generalization; but the sample is small and time-series/multi-relational/multimodal are excluded.

pith-pipeline@v1.3.0-alltime-deepseek · 20322 in / 11018 out tokens · 100685 ms · 2026-08-01T02:38:56.461399+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Data Quality Profiling at Scale with Progressive Sampling: A Benchmark for Data-Centric AI Pipelines." pith.science (2026). https://pith.science/paper/S4IJD2TY

@misc{pith2026260725356,
  author       = {Pith},
  title        = {Pith review of: Data Quality Profiling at Scale with Progressive Sampling: A Benchmark for Data-Centric AI Pipelines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4IJD2TY}},
  note         = {Machine review of arXiv:2607.25356}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Data quality profiling -- computing missing-value rates, duplicate fractions, outlier densities, and functional-dependency violations -- is foundational for data-centric AI pipelines, yet exhaustive scans over millions of rows are prohibitively slow for near-real-time monitoring. Progressive sampling is the standard alternative; the open question is which strategy best preserves profile fidelity at scale. We benchmark nine sampling strategies -- blind (random uniform, geometric, Yamane, cluster) and proxy-guided (Metropolis-Hastings, DAG, stratified by column type or quality score, importance-weighted) -- on three real-world datasets (NYC 311, NYPD arrests, UCI Adult; up to 500K rows), an IoT sensor stream (2.3M rows), two ultra-large real datasets including Ultra-Marathon Running (up to 7.4M rows), and synthetic data scaled to 5x10^6 rows. Contrary to the assumption sharpens estimates, blind representative samplers dominate uniformly. At a 5% budget, random uniform achieves 0.49% mean relative error on NYC 311; DAG-guided MCMC yields 19.5% (approx. 40x worse), and across all real datasets DAG is 11-49x worse (Wilcoxon W=0, p=0.002, n=9 pairs). Cluster sampling matches random uniform (MRE 0.110 vs. 0.111); proxy-guided methods share DAG's failure mode (MRE 0.20-0.35). At scale, random uniform is near-linear (O(N^{0.964})) while DAG is super-linear (O(N^{1.272})), running 28--47x slower on ultra-large data with 6x worse accuracy. The root cause is an IQR proxy mismatch: proxy-guided samplers over-pursue numeric outliers, while quality defects concentrate in categorical columns invisible to the proxy. The actionable finding: representativeness, not domain knowledge, determines sampler quality -- schema-free random uniform or cluster sampling suffices for production-grade quality profiling at scale.

Figures

Figures reproduced from arXiv: 2607.25356 by Laure Berti-Equille.

Figure 1
Figure 1. Figure 1: E1 — Mean relative error per method at b=50% on all four primary datasets. D1=synthetic 100K; D2=NYC 311 (500K); D3=NYPD arrests (500K); D4=UCI Adult census (49K). DAG-MCMC is worst on all real-world datasets (D2–D4); Uniform and Cluster are near zero. 3.4 E4 — Robustness under Error Injection We inject four error types (missing, duplicate, outlier, inconsistency) at rates 1%–30% into D1 and measure mean r… view at source ↗
Figure 2
Figure 2. Figure 2: E3 — Relative error vs. sampling budget b. D2=NYC 311; D3=NYPD arrests; D4=UCI Adult census. dag is budget-invariant at ≈19–27%; RU converges to sub-1% at b=5% on D2 [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) E4 — Robustness: MRE vs. error injection rate (%, D1 synthetic, dag vs. RU, 3 seeds). (b) E5 — Scalability: wall-clock time (s) vs. N rows (D5 SDV synthetic, dag vs. RU, 1–3 seeds; see footnote). (c) E2 — Correlation ablation: MRE vs. Pearson ρ (D1 synthetic, dag vs. MH vs. Gibbs, 5 seeds per ρ). Lower MRE is better [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] 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

34 extracted references · 4 canonical work pages

  1. [1]

    Abedjan, Z., Chu, X., Deng, D., Fernandez, R.C., Ilyas, I.F., Ouzzani, M., Papotti, P., Stonebraker, M., Tang, N.: Detecting data errors: Where are we and what needs to be done? In: VLDB. vol. 9, pp. 993–1004 (2016).https://doi.org/10.14778/ 2994509.2994518

    arXiv 2016

  2. [2]

    VLDB Journal24(4), 557–581 (2015).https://doi.org/10.1007/s00778-015-0389-y

    Abedjan, Z., Golab, L., Naumann, F.: Profiling relational data: A survey. VLDB Journal24(4), 557–581 (2015).https://doi.org/10.1007/s00778-015-0389-y

    work page doi:10.1007/s00778-015-0389-y 2015

  3. [3]

    In: EuroSys

    Agarwal, S., Mozafari, B., Panda, A., Milner, H., Madden, S., Stoica, I.: BlinkDB: Queries with bounded errors and bounded response times on very large data. In: EuroSys. pp. 29–42 (2013).https://doi.org/10.1145/2465351.2465355

    arXiv 2013

  4. [4]

    Knowledge and Information Systems11(2), 191–215 (2007).https: //doi.org/10.1007/s10115-006-0006-x

    Berti-Équille, L.: Data quality awareness: A case study for cost optimal association rule mining. Knowledge and Information Systems11(2), 191–215 (2007).https: //doi.org/10.1007/s10115-006-0006-x

    doi:10.1007/s10115-006-0006-x 2007

  5. [5]

    In: 34th IEEE International Conference on Data Engineering (ICDE)

    Berti-Équille, L., Bonifati, A., Milo, T.: Machine learning to data management: A round trip. In: 34th IEEE International Conference on Data Engineering (ICDE). pp. 1735–1738 (2018).https://doi.org/10.1109/ICDE.2018.00226

    arXiv 2018

  6. [6]

    International Journal of In- formation Quality2(4), 300–323 (2011).https://doi.org/10.1504/IJIQ.2011

    Berti-Équille, L., Comyn-Wattiau, I., Cosquer, M., Kedad, Z., Nugier, S., Peralta, V., Cherfi, S.S.S., Thion-Goasdoué, V.: Assessment and analysis of information quality: A multidimensional model and case studies. International Journal of In- formation Quality2(4), 300–323 (2011).https://doi.org/10.1504/IJIQ.2011. 043780

    work page doi:10.1504/ijiq.2011 2011

  7. [7]

    Proceedings of the VLDB Endowment (PVLDB)11(8), 880–892 (2018)

    Berti-Équille, L., Harmouch, H., Naumann, F., Novelli, N., Thirumuruganathan, S.: Discovery of genuine functional dependencies from relational data with missing values. Proceedings of the VLDB Endowment (PVLDB)11(8), 880–892 (2018). https://doi.org/10.14778/3204028.3204032

    arXiv 2018

  8. [8]

    2103.03098

    Bouthillier, X., Delaunay, P., Bronzi, M., Trofimov, A., Nichyporuk, B., Szeto, J., Sepah, N., Raff, E., Madan, K., Voleti, V., Kahou, S.E., Michalski, V., Serdyuk, D., Arbel, T., Pal, C., Varoquaux, G., Vincent, P.: Accounting for variance in ma- chinelearningbenchmarks.In:MLSys(2021).https://doi.org/10.48550/arXiv. 2103.03098

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.03098 2021

  9. [9]

    Nature Reviews Neuroscience14(5), 365–376 (2013)

    Button, K.S., Ioannidis, J.P.A., Mokrysz, C., Nosek, B.A., Flint, J., Robinson, E.S.J., Munafò, M.R.: Power failure: why small sample size undermines the re- liability of neuroscience. Nature Reviews Neuroscience14(5), 365–376 (2013). https://doi.org/10.1038/nrn3475

    doi:10.1038/nrn3475 2013

  10. [10]

    Foundations and Trends in Databases 4(1–3), 1–294 (2011).https://doi.org/10.1561/1900000004

    Cormode, G., Garofalakis, M., Haas, P.J., Jermaine, C.: Synopses for massive data: Samples, histograms, wavelets, sketches. Foundations and Trends in Databases 4(1–3), 1–294 (2011).https://doi.org/10.1561/1900000004

    doi:10.1561/1900000004 2011

  11. [11]

    In: Information Processing Letters

    Efraimidis, P.S., Spirakis, P.G.: Weighted random sampling with a reservoir. In: Information Processing Letters. vol. 97, pp. 181–185 (2006).https://doi.org/10. 1016/j.ipl.2005.11.003

    2006

  12. [12]

    VLDB Journal31, 1103–1126 (2022).https://doi

    Grafberger, S., Groth, P., Stoyanovich, J., Schelter, S.: Data distribution debugging in machine learning pipelines. VLDB Journal31, 1103–1126 (2022).https://doi. org/10.1007/s00778-021-00726-w

    work page doi:10.1007/s00778-021-00726-w 2022

  13. [13]

    In: SIGMOD

    Haas, P.J., Hellerstein, J.M.: Ripple joins for online aggregation. In: SIGMOD. pp. 287–298 (1999).https://doi.org/10.1145/304181.304208

    arXiv 1999

  14. [14]

    Hellerstein, J.M.: Quantitative data cleaning for large databases. Tech. rep., United Nations Economic Commission for Europe (UNECE) (2008),http://db. cs.berkeley.edu/jmh/papers/cleaning-unece.pdf 28 L. Berti-Équille

    2008

  15. [15]

    In: SIGMOD

    Hellerstein, J.M., Haas, P.J., Wang, H.J.: Online aggregation. In: SIGMOD. pp. 171–182 (1997).https://doi.org/10.1145/253262.253291

    arXiv 1997

  16. [16]

    Journal of the American Statistical Association47(260), 663–685 (1952).https://doi.org/10.1080/01621459.1952.10483446

    Horvitz, D.G., Thompson, D.J.: A generalization of sampling without replacement from a finite universe. Journal of the American Statistical Association47(260), 663–685 (1952).https://doi.org/10.1080/01621459.1952.10483446

    arXiv 1952

  17. [17]

    John, G.H., Langley, P.: Static versus dynamic sampling for data mining. In: KDD. pp. 367–370 (1996),https://cdn.aaai.org/KDD/1996/KDD96-069.pdf

    1996

  18. [18]

    Kaggle: The big dataset of ultra-marathon running.https://www.kaggle. com/datasets/aiaiaidavid/the-big-dataset-of-ultra-marathon-running/ data(2021),https://www.kaggle.com/datasets/aiaiaidavid/ the-big-dataset-of-ultra-marathon-running/data, accessed: May 2026

    2021

  19. [19]

    In: Advances in Neural Information Processing Systems 36 (NeurIPS), Datasets and Benchmarks Track (2023).https://doi.org/10.48550/ arXiv.2207.10062

    Mazumder, M., Banbury, C., Yao, X., Karlaš, B., Rojas, W.G., Diamos, S., Di- amos, G., He, L., Parrish, A., Kirk, H.R., et al.: Dataperf: Benchmarks for data- centric AI development. In: Advances in Neural Information Processing Systems 36 (NeurIPS), Datasets and Benchmarks Track (2023).https://doi.org/10.48550/ arXiv.2207.10062

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.10062 2023

  20. [20]

    Morgan & Claypool Publishers (2010).https://doi.org/10.1007/978-3-031-01835-0

    Naumann, F., Herschel, M.: An Introduction to Duplicate Detection. Morgan & Claypool Publishers (2010).https://doi.org/10.1007/978-3-031-01835-0

    work page doi:10.1007/978-3-031-01835-0 2010

  21. [21]

    Journal of the Royal Statistical Society97(4), 558–625 (1934).https://doi.org/10.2307/2342192

    Neyman,J.:Onthetwodifferentaspectsoftherepresentativemethod:Themethod of stratified sampling and the method of purposive selection. Journal of the Royal Statistical Society97(4), 558–625 (1934).https://doi.org/10.2307/2342192

    doi:10.2307/2342192 1934

  22. [22]

    In: SIGMOD

    Park, Y., Mozafari, B., Sorenson, J., Wang, J.: VerdictDB: Universalizing approx- imate query processing. In: SIGMOD. pp. 1461–1476 (2018).https://doi.org/ 10.1145/3183713.3196905

    arXiv 2018

  23. [23]

    Patel, H., Guttula, S., Mittal, R.S., Manwani, N., Berti-Équille, L., Manatkar, A.: Advances in exploratory data analysis, visualisation and quality for data centric AI systems. In: KDD. pp. 4814–4815 (2022).https://doi.org/10.1145/3534678. 3542604

    doi:10.1145/3534678 2022

  24. [24]

    Provost, F., Jensen, D., Oates, T.: Efficient progressive sampling. In: KDD. pp. 23–32 (1999).https://doi.org/10.1145/312129.312188

    arXiv 1999

  25. [25]

    PVLDB1(1), 797–808 (2008).https://doi.org/10.14778/1453856.1453943

    Ré, C., Suciu, D.: Approximate lineage for probabilistic databases. PVLDB1(1), 797–808 (2008).https://doi.org/10.14778/1453856.1453943

    arXiv 2008

  26. [26]

    In: VLDB

    Rekatsinas, T., Chu, X., Ilyas, I.F., Ré, C.: HoloClean: Holistic data repairs with probabilistic inference. In: VLDB. vol. 10, pp. 1190–1201 (2017).https://doi. org/10.14778/3137628.3137631

    arXiv 2017

  27. [27]

    Proceedings of the VLDB Endow- ment (PVLDB)11(12), 1781–1794 (2018).https://doi.org/10.14778/3229863

    Schelter, S., Lange, D., Schmidt, P., Celikel, M., Biessmann, F., Grafberger, A.: Automating large-scale data quality verification. Proceedings of the VLDB Endow- ment (PVLDB)11(12), 1781–1794 (2018).https://doi.org/10.14778/3229863. 3229867

    doi:10.14778/3229863 2018

  28. [28]

    Journal of the American Statisti- cal Association85(412), 1050–1059 (1990).https://doi.org/10.1080/01621459

    Thompson, S.K.: Adaptive cluster sampling. Journal of the American Statisti- cal Association85(412), 1050–1059 (1990).https://doi.org/10.1080/01621459. 1990.10474975

    arXiv 1990

  29. [29]

    In: Proc

    Tolle, G., Polastre, J., Szewczyk, R., Culler, D., Turner, N., Tu, K., Burgess, S., Dawson, T., Buonadonna, P., Gay, D., Hong, W.: A macroscope in the redwoods. In: Proc. 3rd ACM Conference on Embedded Networked Sensor Systems (SenSys) (2005).https://doi.org/10.1145/1098918.1098925,IntelBerkeleyResearchLab dataset:http://db.csail.mit.edu/labdata/

    arXiv 2005

  30. [30]

    ACM Transactions on Mathemat- ical Software11(1), 37–57 (1985).https://doi.org/10.1145/3147.3165 Data Quality Profiling at Scale with Progressive Sampling 29

    Vitter, J.S.: Random sampling with a reservoir. ACM Transactions on Mathemat- ical Software11(1), 37–57 (1985).https://doi.org/10.1145/3147.3165 Data Quality Profiling at Scale with Progressive Sampling 29

    arXiv 1985

  31. [31]

    VLDB Journal32(4), 791–813 (2023).https://doi.org/10.1007/s00778-022-00775-9

    Whang, S.E., Roh, Y., Song, H., Lee, J.G.: Data collection and quality challenges in deep learning: A data-centric AI perspective. VLDB Journal32(4), 791–813 (2023).https://doi.org/10.1007/s00778-022-00775-9

    doi:10.1007/s00778-022-00775-9 2023

  32. [32]

    Harper and Row, New York, 2nd edn

    Yamane, T.: Statistics: An Introductory Analysis. Harper and Row, New York, 2nd edn. (1967),https://openlibrary.org/isbn/9780060473136

    arXiv 1967

  33. [33]

    ACM Computing Surveys57(5) (2025)

    Zha, D., Bhat, Z.P., Lai, K.H., Yang, F., Jiang, Z., Zhong, S., Hu, X.: Data- centric artificial intelligence: A survey. ACM Computing Surveys57(5) (2025). https://doi.org/10.1145/3711118

    doi:10.1145/3711118 2025

  34. [34]

    In: Confer- ence on Innovative Data Systems Research (CIDR) (2025),https://www.vldb

    Zhu, Y.: Efficient approximate query processing with block sampling. In: Confer- ence on Innovative Data Systems Research (CIDR) (2025),https://www.vldb. org/cidrdb/papers/2025/p3-zhu.pdf

    2025