Kdd workflow

KDD Workflow

Use this as the project-management layer. KDD's calendar is structurally different from single-deadline ML venues: two cycles per year, each with an abstract deadline about one week before the paper deadline (2026 anchors: Jul 24/Jul 31, 2025 and Feb 1/Feb 8, 2026, AoE), a rebuttal window inside review, and a Resubmit outcome that connects the cycles into a pipeline. Plan the pipeline, not a single date.

KDD is an ACM conference, not a journal: no standing editor, leadership rotates per edition (the KDD 2026 committee roster lives on the conference site; individual names 待核实 before citing), and the cost surface is registration plus, under ACM Open, a possible APC for authors at non-participating institutions — with a temporary 65% ACM subsidy announced for 2026. Verify all money facts against current pages.

Cycle-choice decision

Situation	Choose	Reasoning
Evidence complete, paper drafted	Nearest cycle	Waiting adds staleness risk in fast subfields
Core result solid, ablations/scale runs missing	Next cycle	Rebuttal cannot carry links, so missing evidence has no delivery channel later
Prior cycle returned Resubmit	Next cycle, always declared	Addressed resubmissions carry better odds per the CFP; the forum id travels with you
ADS paper, launch too recent to measure	Cycle after the measurement window closes	The desk-check needs quantified post-launch numbers, not projections

Pipeline milestones (per cycle)

• T-10 weeks — evidence lock plan: four-axis experiment plan agreed (kdd-experiments); ADS measurement windows already running.
• T-6 weeks — mechanism freeze: no new components after this point; ablation matrix enumerable.
• T-4 weeks — draft in acmart: writing directly in the two-column format; scale runs queued (cluster time is the real deadline for large-data papers).
• T-2 weeks — artifact freeze: anonymized repository built fresh, cited in the PDF, smoke-tested on a clean machine (kdd-artifact-evaluation).
• T-1 week — abstract deadline: title/abstract/authors/track registered; all co-author OpenReview profiles complete (five-year history, DBLP, ORCID — chasing this is a named person's job).
• T-0 — paper deadline: submission audit (kdd-submission); ADS deployment quantification pointed to; resubmission declarations attached.
• Review window — rebuttal capacity: reserve author-days in advance; per-review responses under a link ban are writing-intensive (kdd-author-response).
• Decision — branch: Accept → e-rights + TAPS + APC status check (kdd-camera-ready); Resubmit → change-summary plan against the review contract; Reject → routing decision (kdd-topic-selection).

The Resubmit loop as a plan

# kdd-pipeline.yaml - one paper, tracked across cycles
paper: streamhive
track: research
cycle_history:
  - cycle: 2026-C1
    outcome: resubmit
    contract:            # concerns named in reviews = next cycle's spec
      - temporal-split rebuild (R2)
      - decay-mechanism ablation (R1)
      - scale claim beyond 10M edges (R3)
  - cycle: 2026-C2
    changes_summary_page: drafted   # mandatory first page of the PDF
    prior_forum_id: declared
    new_evidence: [table5_ablation, 2.1B_run, rebuilt_splits]
owners:
  profiles_gate: <name>       # co-author OpenReview completeness
  artifact_freeze: <name>
  rebuttal_window: <name, name>

The discipline: the review contract is written down the day reviews arrive, and every contract item maps to a change-summary line before the next cycle's abstract deadline.

Failure modes on this calendar

• Treating the abstract deadline as soft: at KDD it gates the submission and the track choice; a placeholder abstract in the wrong track is a real incident.
• Discovering an incomplete co-author profile on paper-deadline day — the one blocker that requires someone else's login to fix.
• Queuing the largest-scale run after the draft: at KDD the scale run is the contribution evidence, not a garnish, and clusters have queues.
• Winning a Resubmit and then re-litigating the reviews instead of executing the contract — the returning AC checks the named concerns first.
• Forgetting that notification and camera-ready dates for the current cycle were never verified (待核实) and promising co-authors a calendar the venue never published.

Vignette: one paper, three cycles, two tracks

A team's trajectory that the pipeline model handles cleanly: Cycle 1, the sampling method goes to Research and draws a Resubmit with three named concerns. Between cycles, the method also ships inside the company's ranking system. Cycle 2 now offers a fork: return the Research paper with the contract executed, or wait one more cycle and submit the deployed system to ADS with a real measurement window. The pipeline answer: do both, sequenced — Research resubmission in Cycle 2 (the contract items are done and Resubmit standing decays with staleness), ADS submission the following cycle once the post-launch window closes. Two papers, two tracks, no dual-submission conflict, each riding the evidence it actually has (kdd-topic-selection governs the fork; kdd-related-work governs the overlap declarations between them).

Capacity planning numbers

Planning heuristics — replace with your team's observed values after one cycle:

• Rebuttal window: one author-day per review for text-only responses; double it if any new run is permitted and needed.
• Anonymized artifact build from a mature internal repo: 2-3 days including the clean-machine smoke test, dominated by de-identification review.
• ADS measurement window: whatever the CFP demands is a floor; credible pre/post or A/B windows are typically weeks, which is why the ADS decision happens a quarter early, not a deadline-month early.
• Profile-gate sweep across N co-authors: an hour of work, a week of waiting on other people — start it at abstract-minus-two-weeks.

Output format

[Pipeline stage] planning / evidence / drafting / submitted-C<k> / rebuttal / decision-branch
[Cycle target] <cycle + abstract/paper dates from current CFP, AoE>
[Critical path] <three tasks that decide readiness>
[Resubmit contract] <items + owners, or N-A>
[Money checks] ACM Open status / APC exposure / registration (待核实 items flagged)
[Owner map] profiles: <who> / artifact: <who> / rebuttal: <who>