Wsdm writing style

WSDM Writing Style

Revise a manuscript into the register WSDM's series page announces for itself: "practical yet principled approaches" to search and mining on the Web and the Social Web. That phrase is an editorial instruction. A WSDM paper earns its slot by being simultaneously deployable-sounding and analytically honest - pure benchmark exercises read as under-motivated, pure war stories read as unscientific. The stylistic moves below operationalize that balance.

Open with the behavioral problem, not the model family

WSDM work starts from something users, platforms, or web data actually do: clicks are position-biased, sessions are short, graphs are power-law sparse, feedback loops entrench popularity, adversaries game ranking. The strongest openings name that behavioral fact and make the technical contribution the response to it.

Weak (model-first):
  "Transformer-based architectures have shown promise in recommendation.
   We propose a novel attention mechanism..."

WSDM-shaped (behavior-first):
  "Users abandon a search session after roughly one unsatisfying page,
   so a ranker trained on clicks observes almost nothing below rank ten.
   We show this truncated feedback systematically distorts <X>, and give
   an estimator that corrects it using only logged data."

The second version does three venue-calibrated things: it grounds the problem in interaction behavior, it implies the data regime (logs, not labels), and it scopes the contribution ("using only logged data") before a reviewer asks.

The self-containment voice

Because WSDM historically offers no rebuttal, style must do the job dialogue does elsewhere. Concretely:

• Answer objections in-line, at the moment they arise. After a surprising result, the next sentence should be the control that rules out the boring explanation - not a hope that reviewers won't think of it.
• Pre-commit scope. Write claim sentences whose quantifiers survive hostile reading: "improves nDCG@10 on three of four public benchmarks under temporal splits" is defensible; "consistently outperforms state of the art" invites a terminal review.
• Show the failure mode. One honest paragraph on where the method breaks (cold users, head queries, adversarial content) converts a reviewer's discovery into your demonstration of judgment.

Sentence-level tells, WSDM edition

Draft tell	Why it hurts here	Repair
"web-scale" with no numbers	Industry reviewers know what scale looks like	Give corpus/user/query counts, even rounded
"significantly better" without a test	Ranking metrics are noisy; the pool knows it	Name the test, unit (query/user), and p-value or CI
"real-world dataset" for MovieLens	Reads as inflation to log-data veterans	Call public benchmarks what they are
"novel framework"	Framework-speak hides the mechanism	Name the one thing that changed and its effect
Unexplained acronym soup in the intro	Single-track audience spans IR, graphs, ads, recsys	Expand on first use; the reader may be from the adjacent subfield

The last row is genuinely WSDM-specific: a single-track community means your talk - and your reviewers - include people from every corner of the venue's scope. Write the introduction for the intelligent non-specialist in the adjacent chair.

The ethics section as prose, not boilerplate

WSDM requires an ethical-considerations discussion and excludes it from the page count. Reviewers can smell a template. Write it like a technical section:

1. Name the specific population and data (whose behavior, collected how, under what terms).
2. Name the sharpest realistic harm of your contribution (a better ranker can amplify popularity bias; a deanonymization-adjacent mining result can be misused; an ad-targeting improvement has manipulation surface).
3. State mitigations you actually implemented (aggregation, differential thresholds, release policy), not aspirations.
4. If the honest answer is "low marginal risk," argue it in two sentences instead of padding.

Compression without loss

The 9-page inclusive budget (2026 anchor; see wsdm-supplementary for the arithmetic) rewards specific habits:

• One idea per paragraph; delete paragraph-final summary sentences that repeat the paragraph.
• Kill the second example everywhere: one well-chosen example plus a precise general statement beats two examples.
• Move mechanics into figures: a pipeline figure with a two-line caption often replaces half a page of process prose.
• Related work as positioning, not survey - three contrast sentences beat fifteen citation-list sentences (method in wsdm-related-work).
• Never compress: metric definitions, split protocols, or the limitation paragraph. These are the sentences the SPC quotes when defending the paper.

Abstract anatomy for this venue

Six sentences carry a WSDM abstract: behavioral fact; consequence for search/mining/recommendation; the mechanism, named; evaluation regime (datasets, split type, test) attached to the main number; the online or deployment sentence if one exists, protocol-tagged; the artifact/limitation sentence. The worked rewrite in ../../resources/worked-examples/01-introduction.md executes exactly this skeleton - compare your abstract against it sentence by sentence, and delete anything that is neither one of the six nor evidence for one.

Revision pass order

1. Claims pass: extract every claim sentence; check each against its evidence and its quantifier.
2. Objection pass: inject in-line answers from the objection inventory (wsdm-author-response).
3. Register pass: apply the tells table; ground openings in behavior.
4. Budget pass: cut to fit using the priority order above.
5. Cold-reader pass: someone from an adjacent WSDM subfield reads the intro and restates the contribution; if their restatement is wrong, rewrite.

Output format

[Register] behavior-first opening: yes / rewritten
[Claim audit] N claims; overquantified: <list with repairs>
[Self-containment] in-line objection answers added at: <locations>
[Tells] instances found -> fixed: <count by type>
[Ethics prose] specific population/harm/mitigation named: yes / boilerplate
[Cold-reader test] contribution restated correctly: yes / no