kdd-submission
GitHub用于审核KDD论文提交就绪状态，涵盖轨道声明、OpenReview资料完整性、8页格式限制、双盲审查及AI披露等关键检查点，防止因格式或流程错误导致拒稿。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill kdd-submission -g -y
SKILL.md
Frontmatter
{
"name": "kdd-submission",
"description": "Use when auditing a KDD submission for readiness in either cycle, covering Research vs ADS track declaration, OpenReview profile completeness for all authors, the ACM sigconf 8-content-page budget, double-blind checks, in-paper artifact references, resubmission declarations, and desk-reject triggers specific to ACM SIGKDD."
}
KDD Submission
Run this audit twice: once when the abstract is registered, once before the full-paper deadline. KDD's abstract and paper deadlines are about a week apart in each cycle (2026 Cycle 1: July 24 / July 31, 2025; Cycle 2: February 1 / February 8, 2026, all AoE), and several failure modes below are only fixable in that gap. Reopen the current track CFP first — every number here is a 2026-cycle anchor, not a permanent rule.
Track and cycle declaration
- One paper, one track: the same work may not go to the Research Track and the Applied
Data Science Track simultaneously. If the team is still arguing about track choice at
submission week, run
kdd-topic-selectionbefore anything else. - Confirm which cycle the deadline belongs to. A team planning around "the KDD deadline" without naming the cycle is planning around the wrong date half the time.
- Resubmissions from a prior cycle's Resubmit decision must be flagged in the form, must carry the previous OpenReview forum id, and must prepend a one-page summary of changes as the first page of the PDF. Omitting the summary page wastes the resubmission advantage.
OpenReview profile gate
KDD 2026 required a complete OpenReview profile for every listed author, not just the submitter: institutional affiliations covering at least the past five years, homepage, DBLP link where prior publications exist, ORCID, advisors, and recent publications. Incomplete co-author profiles are a pre-deadline blocker that no one person can fix alone — chase them at abstract time, not on deadline night.
Format audit
The Research Track submission is 8 content pages, followed by references and an optional appendix, in the double-column ACM format:
% KDD 2026 Research Track recommended setting (verify against the current CFP)
\documentclass[sigconf,anonymous,review]{acmart}
% - sigconf: ACM conference two-column layout
% - anonymous: strips author block for double-blind review
% - review: adds line numbers reviewers cite in their reports
% Do NOT tweak \baselineskip, margins, or font sizes; ACM/TAPS tooling and
% chairs both detect a modified acmart.
- Content beyond page 8 that is not references or appendix is a desk-level risk; move it or cut it, never squeeze it.
- Anonymity must survive the appendix and the referenced repository: author names, affiliations, acknowledgements, grant numbers, institutional dataset names, and git history are all leak surfaces.
- Reference the code/data repository inside the PDF: KDD's rebuttal and discussion phases forbid hyperlinks, so a repository not cited at submission time is invisible to reviewers forever.
- Complete the generative-AI disclosure in the submission form if any such tool touched the text or the work; ACM's authorship policy also governs who may appear as an author.
Desk-reject and severity table
| Trigger | Track affected | Severity | Fixable after deadline? |
|---|---|---|---|
| Missing post-launch performance quantification | ADS | Desk reject per the 2026 CFP | No |
| Content pages over the budget | Both | Desk-level | No |
| Identity leak in PDF, appendix, or repo | Both | Desk-level or chair flag | No |
| Incomplete co-author OpenReview profile | Both | Submission blocked or flagged | Only before deadline |
| Undeclared resubmission of a Resubmit paper | Both | Loses resubmission standing | No |
| Same paper in Research and ADS | Both | Rejection of both | No |
| Repository link absent from PDF | Both | Not fatal, but unrepairable in rebuttal | No |
Final-week sequence
- Freeze the track and cycle declaration; verify the OpenReview group id matches
(e.g., a
Research_Track_Cycle_2group is not the ADS group). - Sweep all co-author profiles against the completeness list above.
- Rebuild the PDF from a clean acmart checkout; confirm 8 content pages and that the appendix starts after the references.
- Anonymize the artifact: fresh repo without history, no institutional paths, README free of names; cite it in the paper body.
- For ADS: point at the exact section and table quantifying post-launch performance — if a reviewer cannot find it in 30 seconds, neither will the desk check.
- Register the abstract early in the week; upload the paper with hours, not minutes, of margin against AoE.
Vignette: an ADS submission that nearly desk-rejected itself
A payments team submits their deployed risk model to the ADS track. The draft's evaluation section leads with offline AUC across four benchmark datasets — strong work, but the post-launch numbers sit in a single sentence inside the conclusion ("since launch, chargebacks decreased"). Under the 2026 ADS rule this reads as unquantified deployment evidence: no metric definition, no window, no baseline period. The audit fix costs one evening, not one experiment:
- Promote the production measurement to its own section with a table: metric definition, measurement window, traffic share, before/after or A/B values.
- Demote two of the four offline benchmarks to the appendix; ADS reviewers weight the online table far above the third and fourth offline dataset.
- Add one paragraph on what the offline metrics failed to predict about online behavior — the lessons-learned register the track exists for.
The general lesson: at submission time, audit findability of the track's gate evidence, not just its existence.
Cross-cycle edge cases
- A paper rejected (not Resubmitted) in Cycle 1 and substantially rewritten may return in Cycle 2 as a fresh submission — but if the overlap with the old forum is obvious, silence is riskier than a voluntary note; check the current CFP's wording on prior-cycle papers before deciding.
- A Resubmit paper whose authors switch tracks (Research → ADS) is effectively a new submission: the Resubmit standing attaches to the reviewed paper, not to the project. Re-read the current rules before assuming the change summary carries over.
- Adding authors between abstract and paper deadlines can trip the profile gate again at the worst moment: every new name needs a complete OpenReview profile too.
- The abstract registered at the abstract deadline should match the final framing; bait-and-switch abstracts (registered as graph mining, submitted as LLM prompting) risk mis-bidding and hostile expert assignment.
Output format
[KDD readiness] Ready / Needs fixes / Not ready
[Track x cycle] Research|ADS x Cycle 1|Cycle 2 (source: current CFP)
[Blocking checks] <profiles/pages/anonymity/track duplication/resubmission flag>
[ADS deployment gate] quantified post-launch evidence: present / absent / N-A (Research)
[Artifact visibility] repo cited in PDF: yes / no
[Fix order] <ordered fixes before the AoE deadline>
版本历史
- 9f86f09 当前 2026-07-19 16:56
