brycewang-stanford KDD Topic Selection
Use this before any writing. KDD routing is a two-stage decision — first whether the work is KDD-shaped at all, then which track — and the second stage is the one teams get wrong, because the Research and ADS tracks want different papers written about the same system. A submission may enter exactly one of them per cycle.
Stage 1: is it KDD-shaped?
KDD rewards contributions where the data regime is the protagonist: scale, drift, heterogeneity, noise, sparsity, graph structure, streams, or the friction of real deployment. Signals:
- The method's value claim references a property of data (works at 10^9 edges, under drift, with 0.1% labels) rather than only a property of models.
- There is a mining/discovery/prediction task with measurable output — not purely a learning-theory statement or a pure systems benchmark.
- Someone outside the authors' subfield could use the result on their data.
Anti-signals: novelty lives entirely in architecture or loss design (ML flagships); the contribution is query processing or storage (VLDB/SIGMOD); the core is retrieval ranking on web corpora (SIGIR/WSDM/WWW may fit better, though overlap is real).
Stage 2: the track fork
The decisive question: what is the strongest evidence you can honestly produce?
| Your strongest honest evidence | Track | The bar you must clear |
|---|---|---|
| New mechanism + ablations + scale curves on (mostly) public data | Research | Delta over the mining lineage, mechanism isolated |
| A system running in production with measurable post-launch outcomes | ADS | Quantified post-launch performance — its absence is a desk reject (2026 CFP); lessons learned expected |
| A dataset/benchmark others will build on, with licenses and baselines | Datasets & Benchmarks | Documentation, provenance, maintenance story |
| Mining/ML advancing a natural-science question | AI for Sciences | Scientific validity alongside method quality |
| A provocative direction without full evidence | Blue Sky Ideas | Argument quality; check the track's own CFP |
Two frequent mis-forks:
- Deployed system submitted to Research because the team fears the ADS bar: it then competes on mechanism novelty it may not have, while its actual strength (deployment evidence) earns no credit.
- Research prototype submitted to ADS with projected or offline-only numbers: desk-rejected under the post-launch quantification rule. "We could deploy it" is a Research-track sentence.
- The blocked-from-deployment exception in the ADS CFP is for documented external blockers (regulation, partner constraints), not for "we ran out of time."
Decision vignette
A team built a fraud-detection model, ran it in shadow mode for two months, and has a
new graph-sampling trick inside it. Three honest papers hide here: (a) the sampling
trick with ablations and public-graph scale curves → Research; (b) after full
launch and a measured A/B window → ADS; (c) the labeled transaction graph, if
releasable → Datasets & Benchmarks. The wrong move is one paper claiming all
three, thinly. Pick the paper whose evidence exists this cycle — the dual-cycle
calendar (kdd-workflow) makes "ADS next cycle, after the measurement window" a
concrete plan rather than a consolation.
Neighbor-venue routing
Route AWAY from KDD when:
- contribution = new learning objective/architecture, data regime incidental
-> NeurIPS / ICML / ICLR
- contribution = mining method, but community/deadline fit favors a sibling
-> ICDM or SDM (same shape, different scale of venue)
- contribution = web/user-behavior/retrieval centric
-> WWW / WSDM / SIGIR / RecSys
- contribution = data management, indexing, query execution
-> VLDB / SIGMOD
- contribution = domain science outcome, method standard
-> domain journal, or KDD's AI-for-Sciences track if the mining is real
If the paper misses at KDD, the reviews tell you which axis failed
(kdd-review-process); a mechanism-novelty objection with solid deployment evidence
often means the right venue was the ADS track all along — a track switch, not a
venue switch.
Fit-signal reference
| Signal in the project | KDD reading |
|---|---|
| Method's win grows with data size or graph density | Core Research fit — scale is the venue's home turf |
| Production system + measurable online outcomes | ADS fit, and only ADS once post-launch numbers exist |
| Contribution is a labeled corpus or benchmark harness | Datasets & Benchmarks, not a Research paper padded with baselines |
| Novel loss/architecture, benchmarks incidental | ML flagship shape; expect "delta over lineage?" pushback at KDD |
| Discovery claim in a natural-science domain | AI for Sciences, or a domain venue if the mining is off-the-shelf |
| Query/index/storage performance is the headline | VLDB/SIGMOD shape |
| The paper's tables would not change if the data were 100x smaller | The scale story is decorative — fix that before routing anywhere |
Timing interacts with routing
Because KDD runs two cycles, track choice is also schedule choice: a Research paper
can target the nearest cycle, while the ADS version of the same project may need to
wait for its measurement window to close (kdd-workflow). When co-authors disagree
on track, the tiebreaker question is: which paper's gate evidence exists today?
Projected evidence loses to existing evidence at this venue every time — the ADS
desk-check makes that literal.
Commit checklist before writing
- One sentence naming the data regime and why it defeats existing methods.
- Track chosen from the evidence table above, and the track's specific bar quoted from the current CFP (track CFPs are separate documents — read the right one).
- The nearest KDD-lineage ancestor identified (
kdd-related-work). - The cycle chosen with evidence lead times honored (
kdd-workflow).
Output format
[KDD-shaped] yes / no -> <redirect venue>
[Track] Research / ADS / Datasets+Benchmarks / AI4Science / BlueSky
[Evidence basis] <the strongest honest evidence available this cycle>
[Desk-check exposure] ADS post-launch quantification: satisfied / fails / N-A
[Regime sentence] <one sentence>
[Next action] <write / measure / wait a cycle / re-route>