Which labs, teams and papers most shaped supervised learning on tabular data — trees, neural networks, foundation models and benchmark papers — with authors, Google Scholar citations and how impact grew over the last ten years, plus what practitioner forums say, and recent papers included.
This review curates 19 pivotal papers (from 310 extracted candidate rows spanning 2016–2026) and 29 citation-result rows, keeping only citation counts backed by a Google Scholar page (snapshot 2026-10-28, approximate). It is a curated impact review drawn from scans of 1,900 candidate web results for papers and 1,410 for forums — not an enumeration of every tabular-ML paper. Google Scholar did not expose machine-readable annual per-paper counts, so no yearly citation curves are drawn; velocity below is current citations divided by years since publication, a rough rate only.
Bubble area = current Google Scholar citations (snapshot 2026-10-28, approximate; shown only where a Google Scholar URL backs the count). Small fixed dots = influential papers without a verified Google Scholar count — recent work is kept, not ranked by citations. Hover or tap a bubble for authors, lab and scope.
| Year | Family | Paper | Authors | Lab / team | Venue | Benchmark scope | GS citations | Source |
|---|
Tianqi Chen and Carlos Guestrin’s 2016 KDD system paper remains the single most cited work in tabular ML (~95,796 on Google Scholar) and the baseline every later paper measures against.
Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye and Tie-Yan Liu built the fast histogram-based GBDT (~31,551 citations); author-profile mirrors vary around 30.6k–31.6k.
Liudmila Prokhorenkova’s team delivered CatBoost (~11,822), while Gorishniy, Rubachev, Khrulkov and Babenko produced FT-Transformer (“Revisiting…”, ~2,707), TabM and the TabReD industry benchmark.
Sercan Arik and Tomas Pfister’s attentive, interpretable architecture (~3,930) made deep tabular learning mainstream in industry toolkits.
Hollmann, Müller, Purucker and Hutter created TabPFN (ICLR 2023) and its Nature 2025 successor (~1,994), then co-built TabArena, TabPrep and Beyond IID — the engine room of the foundation-model wave.
Grinsztajn, Oyallon and Varoquaux wrote the defining trees-vs-DL benchmark; Qu, Holzmüller, Varoquaux and Le Morvan scaled in-context learning (TabICL, TabICLv2); Layer 6 AI (Toronto) contributed LoCalPFN and TabDPT; the TabZilla team spans Abacus.AI, NYU, IIT Bombay, Caltech and more.
Trees hold typical medium data. Grinsztajn, Oyallon and Varoquaux (NeurIPS 2022, 45 datasets) find tree-based models remain state of the art on medium-sized data (~10k samples), tracing it to inductive-bias differences. TabZilla (NeurIPS 2023; 19 algorithms, 176 datasets, 538,650 models trained) concludes the NN-vs-GBDT debate is overemphasized: light GBDT tuning often matters more than the family choice.
Foundation models win defined small-data regimes. The Nature TabPFN paper reports dominance over tuned GBDTs on 29 classification and 28 regression datasets with up to 10,000 samples and 500 features — a stated scope, not all of tabular ML. One 2024 survey of 68 datasets argues a broader paradigm shift toward deep methods.
2026 work re-draws the boundary. Beyond IID (142 datasets, 11 models) finds foundation models excel on tiny-to-medium IID data while trees and deep nets still dominate non-IID, large and high-dimensional tasks; TabReD shows many academic gains fade on 8 industry-grade, temporally split datasets; TabArena (51 datasets, 16 models, ~25M model instances) now tracks the leaderboard continuously.
A small, diverse selection from 312 extracted forum rows. These are recurring practitioner themes — anecdote, not benchmark evidence: GBDTs are easier and often stronger by default; NNs earn their keep for multimodal or streaming settings; foundation models attract interest for fast small-data inference.
Centrally relevant recent methods and benchmarks from the extracted rows; citation column left blank where no verified Google Scholar count exists.
Method: curated impact review of supervised tabular machine learning, 2016–2026. Inputs: 310 extracted candidate paper rows (deduplicated by normalized title to 19 curated pivotal papers), 29 citation-result rows (only Google Scholar-hosted rows used for counts, snapshot 2026-10-28, approximate — mirrors differ, e.g. XGBoost 93,928 vs 95,796), 5 fetched Google Scholar paper pages for author verification, and 312 forum rows (8 shown). Research scanned 1,900 candidate web results for papers and 1,410 for forums. Citation velocity = current citations ÷ years since publication, a rough rate; no annual citation curves were available, so none are drawn. Domain applications, duplicates and generic preprocessing papers were cut for space.