Data-science and machine-learning competitions hosted on platforms like Kaggle, or run as an official track of a research venue such as the NeurIPS Competition Track or ACM SIGKDD’s annual KDD Cup, routinely generate two structurally different kinds of paper: one written by the people who ran the competition, and one (or several) written by the teams who won it. Because these are not standard single-lab manuscripts, standard authorship guidance from ICMJE or a single journal’s policy doesn’t map onto them cleanly, and the venues themselves say relatively little about it directly. This guide lays out what actually governs credit in practice, and where research administrators and participants are left to apply general authorship principles themselves.
Two publication tracks, two author lists
Almost every academically affiliated ML/data-science competition produces up to two distinct papers, with distinct, non-overlapping author lists:
- The organizer / dataset-overview paper. Written by the organizing committee — the people who designed the task, assembled and curated the dataset, defined the evaluation metric, and ran the leaderboard. This paper describes the competition itself: motivation, data collection and cleaning, task formulation, baseline results, and (after the competition closes) a summary of what approaches did well. Real examples of this genre include the annual “KDD Cup Report” papers (for instance, the KDD Cup 2019 organizing-committee report) and overview papers for large Kaggle-hosted challenges written by the academic or industry group that sponsored the competition, such as the ASHRAE Great Energy Predictor III competition overview paper.
- The winning/participant solution paper. Written by a competing team — usually, though not always, a team that placed at or near the top of the leaderboard — describing their modeling approach, feature engineering, ensembling strategy, and what did or didn’t work. These are typically authored solely by the team’s own members, published independently of the organizers’ paper (on arXiv, in an affiliated workshop proceedings volume, or as a competition-track paper at the host conference). A representative example: the winning solution to KDD Cup 2016 was published as its own arXiv paper, authored entirely by the winning team, separately from that year’s official organizers’ report.
These two papers rarely share authors. An organizer is not automatically a co-author on a winning team’s solution paper merely by having designed the task or hosted the leaderboard, and a winning participant is not automatically a co-author on the organizers’ dataset/overview paper merely by having competed. The same logic CASRAI’s broader authorship content applies to data paper authorship vs. data contribution and to citation of AI models and training datasets generally holds here: providing infrastructure, data, or a benchmark task is usually a citation relationship, not an authorship one, unless a specific individual also did substantive intellectual work on the paper being written.
Do competition organizers get co-authorship on a winning team’s paper?
As a general rule, no — not by default. The convention that has emerged across these venues mirrors ordinary academic authorship norms rather than inventing a competition-specific rule:
- If an organizer contributed no more than hosting the platform, defining the task, and releasing the dataset, the appropriate form of credit in the winning team’s paper is a citation to the organizers’ overview/dataset paper (or to the competition itself, if no paper exists yet) — not a byline. This mirrors ordinary data-citation practice: using someone’s dataset or benchmark obligates you to cite it, not to add its creator as a co-author.
- If an organizer (or a member of the organizing committee) genuinely collaborated on the specific solution being written up — for example, by co-developing the modeling approach with the team after the competition closed — ordinary authorship criteria apply and co-authorship can be appropriate. The test is the same one ICMJE applies anywhere else: substantial intellectual contribution to the specific work being published, not proximity to the competition.
- Where a prize sponsor or platform requires winning teams to submit a public methodology write-up as a condition of receiving a prize (a common clause in Kaggle competition rules), that write-up is a deliverable to the organizers, not evidence that organizers co-authored it.
Do winning teams get listed as authors on the organizers’ report?
Typically not as co-authors either, though they are very often credited by name. Organizer/overview papers commonly acknowledge or tabulate top-placing teams — team names, member names, and affiliations, sometimes alongside a short summary of each team’s approach — in a results table, leaderboard summary, or dedicated “top solutions” section, without adding those individuals to the paper’s byline. This is analogous to how a large collaborative study credits contributors who do not meet full authorship criteria in an acknowledgments section rather than the author list (see CASRAI’s guide to authorship in large collaborations for the parallel case in physics megacollaborations, where thousands of contributors are handled by group authorship rather than individual bylines — competition papers use a lighter-weight version of the same separation between “meaningfully credited” and “formally an author”).
Venue-specific notes
NeurIPS Competition Track
NeurIPS runs an official Competition Track with its own submission and review process, distinct from the main track. Two details are directly relevant to authorship and credit:
- Competition proposals are reviewed single-blind — organizer identities are disclosed on submission, not anonymized, unlike the double-blind review NeurIPS uses for its main research track.
- Accepted competitions have a defined path for publishing a post-competition paper after the event: either a submission to the NeurIPS Datasets and Benchmarks / Evaluations track (standard peer review, no guaranteed acceptance), or publication in a dedicated PMLR (Proceedings of Machine Learning Research) volume for NeurIPS competitions, which carries a lighter-weight review with guaranteed acceptance for accepted competitions. These PMLR competition volumes have historically included the organizers’ own overview paper as one entry, with separate, individually authored short papers from top-placing teams describing their winning methods as additional entries in the same volume — again, distinct author lists within one proceedings volume rather than a merged byline.
KDD Cup
KDD Cup, run under ACM SIGKDD, follows the organizer-report-plus-independent-solution-paper pattern described above. The organizing committee (which varies year to year and is typically drawn from the academic and industry co-chairs running that year’s challenge) publishes a report summarizing the competition, data, tracks, and results; winning teams frequently publish their own solution papers separately, on arXiv or in venue proceedings, under their own authorship, citing the official KDD Cup report or dataset description rather than including organizers as co-authors.
Kaggle
Kaggle itself does not publish academic papers — it is a hosting platform, and most Kaggle competitions never produce a formal paper from either side. Where a paper does exist, it usually comes from one of two sources: an academic or industry group that used Kaggle as the hosting mechanism for a competition tied to a specific venue (in which case that group, not Kaggle Inc., is the “organizer” for authorship purposes, and typically publishes its own dataset/overview paper independently of Kaggle), or a participating team choosing to write up its own approach as a technical report or arXiv preprint after the competition closes, entirely under that team’s own authorship. Many Kaggle competitions’ official rules also require top-ranking, prize-eligible teams to submit a written methodology description as a condition of receiving a prize; this is a platform/sponsor deliverable and forum-published write-up, not itself a peer-reviewed authored paper, though teams frequently expand it into one afterward.
How is author order decided within a competing team?
None of the three venues above prescribe an author-order convention for team solution papers — this is left entirely to the team. In practice, competition teams tend to fall back on the same general approaches used in any small collaborative research group, most commonly:
- Contribution-based ordering, informally negotiated among teammates, often placing whoever did the most modeling/write-up work first and the team lead or most senior member last — the same first-author/senior-author convention covered in CASRAI’s guide to CRediT roles and shared credit.
- Platform-role ordering, where the person who submitted on the platform, merged the team, or served as team captain during the competition is listed first, independent of who wrote the most code — a convention specific to how these teams actually form (often ad hoc, sometimes across institutions or countries, coordinated entirely through the platform’s discussion forum rather than a shared employer).
- Alphabetical order, sometimes used by teams that want to sidestep a ranking dispute entirely, occasionally noted explicitly as such in the paper (an “authors listed alphabetically” note) rather than implying seniority.
- Explicit equal-contribution flags, footnoting two or more names as having contributed equally when the team does not want to imply one member outranked another — standard academic practice, not specific to competitions.
Because competition teams often form across institutions, countries, and even employer/non-employer lines — a paid data scientist and a student teaming up purely through the platform, for example — there is no single institutional authorship policy governing the team the way there would be for a single lab’s manuscript. CASRAI’s general guidance is the same one that applies to any ad hoc, cross-institutional collaboration: agree on author order and on what “counts” as a contribution before the paper is drafted, not after results come in, and apply the same substantial-contribution test ICMJE uses elsewhere — actually working on modeling, analysis, or the write-up — rather than defaulting to “everyone on the leaderboard team” if some members only contributed compute or moral support.
Does CRediT apply to competition papers?
None of Kaggle, NeurIPS, or KDD Cup currently require a CRediT (Contributor Roles Taxonomy, formalized as ANSI/NISO Z39.104-2022) statement as a condition of competition-track or post-competition paper submission — CRediT adoption is driven by journals and funders, not by these ML competition venues. That said, CRediT is a useful voluntary framework for competition teams to apply, particularly for larger teams (five or more members is common in top Kaggle placements) where a flat byline understates who actually did what. A team can add its own contribution statement using CRediT’s fourteen roles — for example, distinguishing who ran the modeling and hyperparameter search (Software, Formal Analysis) from who wrote the paper (Writing – Original Draft) from who supplied compute or led the team (Resources, Project Administration) — even though no venue currently mandates it. If a competition write-up is later expanded and submitted to a journal that does require CRediT, working out those roles at write-up time saves re-litigating the question later. See CASRAI’s CRediT author statement samples for the exact phrasing conventions.
Practical guidance for teams and research administrators
- Cite, don’t co-author, the organizers unless a specific organizer did substantive intellectual work on your particular solution paper — hosting the leaderboard and releasing the dataset is a citation obligation, not an authorship one.
- Settle author order before drafting, using an explicit standard (contribution, platform role, or alphabetical) and documenting it, especially for cross-institutional or cross-border teams with no shared authorship policy to default to.
- Check the competition’s own rules for a mandatory write-up clause before assuming a forum post satisfies your institution’s own publication/documentation requirements — a prize-eligibility write-up and a citable academic paper are not automatically the same document.
- If your institution requires researchers to log all outputs (for tenure files, grant reporting, or an institutional CRIS), treat a competition solution paper the same as any other paper: verify author order was agreed, not assumed, and confirm the organizer/dataset citation is present and correctly attributed.
- Use CRediT even where it isn’t required for larger competition teams, so the contribution record survives if the write-up is later expanded into a journal submission.
Frequently asked questions
Does hosting a Kaggle competition entitle the host to co-authorship on winning teams’ papers?
No. Hosting a competition, releasing a dataset, and running a leaderboard is normally credited through citation of the organizers’ own paper or the competition itself, not through co-authorship on papers the organizers didn’t substantively help write.
Is there an official authorship policy for NeurIPS Competition Track papers?
NeurIPS specifies review mechanics for the Competition Track itself (single-blind review, organizer identities disclosed) and two publication pathways for post-competition papers (a Datasets and Benchmarks/Evaluations track submission, or a dedicated PMLR volume with lightweight review), but it does not publish a competition-specific authorship-eligibility policy beyond its general Code of Conduct — ordinary substantial-contribution norms apply to who is listed as an author.
Do all Kaggle competitions produce an academic paper?
No. Most Kaggle competitions never produce a formal paper from either the host or the participants; a paper typically only appears when an academic or industry group used Kaggle to host a challenge tied to a venue, or when a team chooses to write up and publish its own solution independently.
Should a competition team use CRediT for its solution paper?
It isn’t required by any of these venues, but it’s good practice for larger teams, and it makes later expansion into a journal submission — where CRediT may be required — much easier, since the contribution breakdown is already documented.
