Pith. sign in

REVIEW 3 major objections 5 minor 19 references

SALT: Sales Autocompletion Linked Business Tables Dataset

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A linked ERP sales dataset for table representation learning, with a field autocompletion task.

desk verdict SALT is a genuinely useful linked-ERP dataset release with a clearly documented schema and task; the main caveat is that filtering to fully processed orders may undercut the autocompletion realism, though it is addressable. read the letter →

arxiv 2501.03413 v1 pith:JAXZXY7E submitted 2025-01-06 cs.LG cs.AIcs.DB

Tassilo Klein , Clemens Biehl , Margarida Costa , Andre Sres , Jonas Kolk , Johannes Hoffart This is my paper · ORCID
classification cs.LGcs.AIcs.DB
keywords linkedbusinesstablesERPdatasetsalesautocompletiontablerepresentationlearningmulti-tabledatatabularbenchmarksdriftenterprise
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces SALT, a curated dataset sourced from a real Enterprise Resource Planning (ERP) system, designed to support research on table representation learning for linked business tables. The authors argue that existing table corpora come from web scraping and do not capture the complexity of multi-table enterprise data, and that SALT fills this gap by offering four interconnected tables covering sales documents, items, customers, and addresses. The dataset is accompanied by a sales autocompletion task: predicting eight target fields that are typically populated late in the sales order creation process. If the dataset is representative, it gives researchers a public, anonymized benchmark for studying how tabular models handle real-world linked business data, where class imbalance, high cardinality, noise, and data drift are the norm.

What carries the argument

The key object is the dataset itself: four tables (I_SalesDocument, I_SalesDocumentItem, I_Customer, I_AddrOrgNamePostalAddress) linked by primary and foreign keys, which are filtered to sales orders only, anonymized through encryption of categorical variables, and joined into a flat table with one row per sales order item. The task is multiclass classification on seven critical sales-order fields (SalesOffice, SalesGroup, CustomerPaymentTerms, ShippingCondition, ShippingPoint, Plant, and IncotermsClassification), which are marked as target fields because they are often populated later in the order-creation workflow. This setup lets a single model trained on the joined table directly address the business problem of sales autocompletion.

What would settle it

Take a raw, unfiltered ERP sales-order snapshot from the same system and period, and compare the value distributions and target-field missing rates with SALT; if the excluded incomplete orders are disproportionately the ones lacking the target fields, then SALT under-represents the autocompletion scenario and the claim that it reflects actual enterprise business contexts loses support.

Watch

Extended reading notes

Core claim

The central claim is that SALT constitutes a realistic, multi-table enterprise dataset: it contains 500,908 sales orders, 2,319,944 sales order items, 139,611 customers, and 1,788,887 addresses, spanning transactions from January 2018 to December 2020, with temporal splits for validation and test. The paper further claims that the dataset demonstrates the characteristics of actual enterprise systems, including skewed class distributions, long-tailed categorical fields, input noise from manual data entry, and potential data drift over time. Baseline experiments show that most current tabular models handle the prediction tasks reasonably well, with CARTE achieving the best average mean reciprocal rank of 0.79, followed by AutoGluon at 0.77 and XGBoost at 0.76, while majority-class baselines perform far worse on most targets.

Load-bearing premise

The filtering and anonymization pipeline preserves the distributional properties of real ERP data, including the missing-value patterns in the target fields that the autocompletion task is designed to predict.

Editorial extensions

If this is right

  • Researchers gain a public, anonymized multi-table ERP dataset that can serve as a benchmark for table representation learning on linked business data, complementing web-scraped corpora like WebTables and GitTables.
  • Baseline results indicate that gradient-boosted trees and AutoML pipelines already reach high accuracy on some targets (e.g., Sales Office near 0.99 MRR), while high-cardinality fields like Sales Group remain challenging, so the dataset can separate easy saturation from hard prediction.
  • The temporal splits allow evaluation under realistic data drift, where category definitions may shift over time, providing a testbed for drift-robust tabular models.
  • The flat-table format with repeated header-level targets enables multi-task learning where a single model predicts all target fields simultaneously, as the paper explicitly designed for.
  • Because the raw table structure is also released, methods that exploit foreign-key linkages natively, such as graph-based approaches, can be compared against flat-table baselines.

Reading between the lines

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

  • A likely follow-up is to benchmark table foundation models on the original four-table schema rather than the flat join, since the paper's flattening discards the explicit relational structure that distinguishes SALT from web-scraped corpora.
  • The exclusion of incomplete sales orders may remove exactly the cases where autocompletion is most needed, so the dataset's difficulty could be understated compared with live ERP conditions.
  • The reported data drift and 2018–2020 time span make SALT a natural candidate for studying concept drift in enterprise tabular classification, a direction the paper mentions but does not explore.
  • The encryption-based anonymization preserves value cardinalities but not semantic meaning, so models trained on SALT cannot directly transfer to other ERP systems without re-encoding; this suggests the dataset is best used for method development rather than for pretraining deployable models.
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 SALT, a publicly released dataset of four linked tables (sales documents, sales document items, customers, and addresses) extracted from a single enterprise ERP system, containing about 2.3 million sales-order-item rows spanning 2018–2020. The dataset is designed for table representation learning and is paired with a sales-autocompletion task: predict eight target fields (e.g., sales office, plant, shipping point, incoterms) that are typically filled in during later stages of sales-order creation. The authors describe the schema, statistics, and an SQL join procedure, and they report baseline evaluations using random/majority classifiers, XGBoost, LightGBM, CatBoost, CARTE, AutoGluon, and GraphSAGE, measured by mean reciprocal rank (MRR). The central claim is that SALT constitutes a realistic, anonymized, multi-table enterprise dataset with a well-defined task and temporal splits.

Significance. If the dataset is as described, it fills a real gap: most existing table corpora are web-scraped and lack the inter-table foreign-key structure and business semantics of operational ERP data. The paper ships a clear schema, a documented join query, per-field statistics, and baseline code, which are valuable for benchmarking tabular and relational learning methods. The inclusion of a real-world enterprise source and a concrete autocompletion task with temporal splits makes SALT potentially useful for the table-representation-learning community. However, the value of the contribution depends on the dataset faithfully representing the enterprise context in which autocompletion operates, and on the baseline comparisons being statistically meaningful.

major comments (3)
  1. [Sec. A.4 and Sec. 2 (Task)] The filtering step in Sec. A.4 keeps only fully processed sales orders ('sales orders which have gone through the entire business process'), and Sec. 2 motivates the task as predicting fields that are 'typically missing in sales orders' during order creation. Because the evaluation set contains only completed orders, the model never sees the partial-order state in which autocompletion would actually be used, and the target distributions are conditioned on final completion status. This is a load-bearing mismatch: the claim in Sec. 4 that SALT 'demonstrat[es] the characteristics of data within actual enterprise systems' is not established for the autocompletion use case. Please address this by (a) providing evidence that incomplete orders follow the same conditional target distributions, (b) including an analysis or separate split on incomplete orders if available, or (c) re-framing the task as end-of-process field prediction and explicitly documenting the limitation.
  2. [Sec. 3 (Experiments & Results), Table 2] All baseline results in Table 2 are reported as point estimates from a single run, with no error bars, standard deviations, or multiple seeds. Given the small margins separating the top methods (CARTE average MRR 0.79, AutoGluon 0.77, XGBoost 0.76) and the strong class imbalance noted in Sec. 2, the reported ranking may be within run-to-run variation. Please report means and standard deviations across at least five independent seeded runs for every method, or otherwise quantify the uncertainty of the comparisons. This is necessary to support the claim that CARTE performs best on SALT with a margin of +0.02.
  3. [Sec. A.1 (Privacy & Anonymization)] The statement that the privacy-sanitization protocol 'was designed to preclude any likelihood of data distribution distortion or introduction of bias' is an assertion rather than a measured property. Deterministic encryption or hashing can in principle alter cardinalities, introduce collisions, or shift value distributions, and the paper provides no before/after comparison or detailed description of the encryption scheme. Because the representativeness of SALT is a central claim, please provide empirical checks (e.g., distributional comparisons of key fields before and after anonymization) or, at minimum, a precise account of the type of encryption used and a discussion of its known properties and limitations.
minor comments (5)
  1. [Sec. 2 (Task)] The text says the model performs 'multiclass classification on seven critical variables,' but Table 2 lists eight target tasks and the bullet list contains eight field names (the two Incoterms fields are counted separately in the table). Please reconcile the count or clarify that the two Incoterms fields are grouped as one 'variable.'
  2. [Sec. A.3] The text states 'Table 8 provides statistics on the target fields' and then 'Table 8 provides detailed statistics of the table fields'; the second reference should presumably be to a different table (Table 9?), and the manuscript's actual Table 8 is titled 'Field statistics.' Please fix the table numbering and cross-references.
  3. [Table 1 caption] The caption reads 'Atomatic data type'; this should be 'Atomic data type.'
  4. [Sec. 3 (Experiments & Results)] The paper states that the validation set was used for early stopping and that no hyperparameter tuning was performed, but it does not specify how early stopping was implemented (e.g., monitored metric, patience) or whether the validation set was also used for any model selection. Please clarify this procedure.
  5. [Sec. 2 and Table 7] The paper reports missing-value percentages for target fields (up to 1.09% for CustomerPaymentTerms), but it does not describe how rows with missing target labels are treated in training and evaluation. Please state whether such rows are dropped, imputed, or handled as a special class.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SALT is an externally benchmarked dataset release whose central claims do not reduce to their inputs.

full rationale

The paper's load-bearing content is a dataset release plus baseline evaluation, not a derivation. The eight target fields (SalesOffice, SalesGroup, CustomerPaymentTerms, ShippingCondition, ShippingPoint, Plant, and the two Incoterms fields) are selected as fields typically populated later in sales-order creation (Sec. 2, Appendix A.2), and the prediction targets are not defined by any fitted parameter or by the models evaluated. Baseline methods (XGBoost, LightGBM, CatBoost, CARTE, AutoGluon, GraphSAGE) are external implementations evaluated on a fixed train/validation/test temporal split; no result is "predicted" by reinserting a quantity that was used to construct the targets. The only arguable weakness is that Appendix A.4 restricts the data to fully processed orders, which may affect how representative the benchmark is of real-time autocompletion; however, that is a data-quality and external-validity concern, not a circularity of the kind where the conclusion is equivalent to an input by construction. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no renamed empirical pattern presented as a derivation. Accordingly, the paper is self-contained against external benchmarks and merits a circularity score of 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted to data; the paper releases a dataset and evaluates external baselines. The central claim rests on domain assumptions about representativeness (anonymization/filtering preserve distributions), task validity (targets are meaningfully optional), and the temporal split. No new theoretical entities are introduced.

assumptions (3)
  • domain assumption The anonymization and filtering pipeline preserves the statistical distribution of the original ERP data (no bias introduced).
    Invoked in Sec. A.1 ('preclude any likelihood of data distribution distortion or introduction of bias') and Sec. A.4 (filtering to fully processed orders). If false, the dataset is not representative of real enterprise data.
  • domain assumption The eight hand-selected target fields correspond to fields that are 'typically missing' or optional in sales orders and are learnable from the 21 input fields.
    Stated in Sec. 2 Task ('target variables are not always maintained; they are optional and may not be filled out'). If targets are trivially derivable from inputs or not actually optional, the task framing is invalid.
  • domain assumption The temporal split (validation from 2020-02-01, test from 2020-07-01) is appropriate for evaluating generalization under the stated data drift.
    Introduced in Sec. 2 Data Insights; the paper notes data drift may impact temporal splits but does not quantify drift or provide an alternative split analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SALT: Sales Autocompletion Linked Business Tables Dataset." pith.science (2026). https://pith.science/paper/JAXZXY7E

@misc{pith2026250103413,
  author       = {Pith},
  title        = {Pith review of: SALT: Sales Autocompletion Linked Business Tables Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JAXZXY7E}},
  note         = {Machine review of arXiv:2501.03413}
}
read the original abstract

Foundation models, particularly those that incorporate Transformer architectures, have demonstrated exceptional performance in domains such as natural language processing and image processing. Adapting these models to structured data, like tables, however, introduces significant challenges. These difficulties are even more pronounced when addressing multi-table data linked via foreign key, which is prevalent in the enterprise realm and crucial for empowering business use cases. Despite its substantial impact, research focusing on such linked business tables within enterprise settings remains a significantly important yet underexplored domain. To address this, we introduce a curated dataset sourced from an Enterprise Resource Planning (ERP) system, featuring extensive linked tables. This dataset is specifically designed to support research endeavors in table representation learning. By providing access to authentic enterprise data, our goal is to potentially enhance the effectiveness and applicability of models for real-world business contexts.

Figures

Figures reproduced from arXiv: 2501.03413 by the authors.

Figure 1
Figure 1. Dataset Schemas. Schemas for the four tables constituting the SALT dataset. Primary keys are highlighted in bold letters and with a key symbol. Foreign keys interconnecting tables. structured tabular applications presents difficulties (Hulsebos et al., 2023; Van Breugel and Van Der Schaar, 2024), particularly in enterprise settings where data privacy, confidentiality, and commer￾cial interests restrict data access. … view at source ↗
Figure 2
Figure 2. Screenshots of SAP S/4HANA Sales Order User Interface [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 8 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Jan-Micha Bodensohn, Ulf Brackmann, Liane Vogel, Matthias Urban, Anupam Sanghi, and Carsten Binnig. 2024. Llms for data engineering on enterprise data. In Joint proceedings of workshops at the 50th International Conference on Very Large Data Bases (VLDB 2024), Guangzhou, China, August 26 - August 30, 2023, VLDBW 2024, Tabular Data Analysis Workshop Proceedings

    work page 2024

  4. [4]

    Cafarella, Alon Halevy, Daisy Zhe Wang, Eugene Wu, and Yang Zhang

    Michael J. Cafarella, Alon Halevy, Daisy Zhe Wang, Eugene Wu, and Yang Zhang. 2008. https://doi.org/10.14778/1453856.1453916 Webtables: exploring the power of tables on the web . Proc. VLDB Endow., 1(1):538–549

    arXiv 2008

  5. [5]

    Tianqi Chen and Carlos Guestrin. 2016. https://doi.org/10.1145/2939672.2939785 XGBoost : A scalable tree boosting system . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, pages 785--794, New York, NY, USA. ACM

    arXiv 2016

  6. [6]

    Xiang Deng, Huan Sun, Alyssa Lees, You Wu, and Cong Yu. 2020. https://doi.org/10.14778/3430915.3430921 Turl: table understanding through representation learning . Proc. VLDB Endow., 14(3):307–319

    arXiv 2020

  7. [7]

    Yuhao Deng, Chengliang Chai, Lei Cao, Qin Yuan, Siyuan Chen, Yanrui Yu, Zhaoze Sun, Junyi Wang, Jiajun Li, Ziqi Cao, Kaisen Jin, Chi Zhang, Yuqing Jiang, Yuanfang Zhang, Yuping Wang, Ye Yuan, Guoren Wang, and Nan Tang. 2024. https://doi.org/10.14778/3659437.3659448 Lakebench: A benchmark for discovering joinable and unionable tables in data lakes . Proc. ...

    arXiv 2024

  8. [8]

    Gus Eggert, Kevin Huo, Mike Biven, and Justin Waugh. 2023. http://arxiv.org/abs/2310.07875 Tablib: A dataset of 627m tables with context

    arXiv 2023

Show all 19 references
  1. [9]

    Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. 2020. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505

    2020 arXiv

  2. [10]

    Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson, Rex Ying, Jiaxuan You, and Jure Leskovec. 2024. https://proceedings.mlr.press/v235/fey24a.html Position: Relational deep learning - graph representation learning on relational databases . ...

    2024

  3. [11]

    Leo Grinsztajn, Edouard Oyallon, and Gael Varoquaux. 2022. https://openreview.net/forum?id=Fp7__phQszn Why do tree-based models still outperform deep learning on typical tabular data? In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2022

  4. [12]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 1025–1035, Red Hook, NY, USA. Curran Associates Inc

    2017

  5. [13]

    Madelon Hulsebos, C agatay Demiralp, and Paul Groth. 2023. Gittables: A large-scale corpus of relational tables. Proceedings of the ACM on Management of Data, 1(1):1--17

    2023

  6. [14]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. Lightgbm: a highly efficient gradient boosting decision tree. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 314...

    2017

  7. [15]

    Myung Jun Kim, Leo Grinsztajn, and Gael Varoquaux. 2024. Carte: Pretraining and transfer for tabular learning. In Forty-first International Conference on Machine Learning

    2024

  8. [16]

    Oliver Lehmberg, Dominique Ritze, Robert Meusel, and Christian Bizer. 2016. https://doi.org/10.1145/2872518.2889386 A large public corpus of web tables containing time and context metadata . In Proceedings of the 25th International Conference on World Wide Web, WWW 2016, Montr...

    2016

  9. [17]

    Jan Motl and Oliver Schulte. 2024. http://arxiv.org/abs/1511.03086 The ctu prague relational learning repository

    2024 arXiv

  10. [18]

    Liudmila Ostroumova Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. 2018. http://dblp.uni-trier.de/db/conf/nips/nips2018.html#ProkhorenkovaGV18 Catboost: unbiased boosting with categorical features. In NeurIPS, pages 6639--6649

    2018

  11. [19]

    Boris Van Breugel and Mihaela Van Der Schaar. 2024. https://proceedings.mlr.press/v235/van-breugel24a.html Position: Why tabular foundation models should be a research priority . In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings...

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.