recsys-writing-style

RecSys Writing Style

Use this when revising the main paper. RecSys papers need a compact statement of why a

recommendation result matters and enough evaluation detail to survive a reproducibility-minded

reviewer.

Revision rules

• Put the recommendation contribution on the first page: problem, gap, method, offline evidence,

and the offline-to-deployment bridge.

• Make the evaluation protocol explicit early: the split (temporal vs random), whether metrics are

full-ranking or sampled, and that baselines are tuned under an equal budget.

• Pair every empirical claim with a table, an ablation, or an off-policy/A-B result — not with a

superlative.

• Use the 8-page body for core logic; move tuning grids, extra datasets, and per-dataset

breakdowns to the appendix (which counts inside the budget) without making the body unreadable.

• Avoid leaderboard framing ("state of the art," "outperforms all baselines") when the gain is

small, the variance unreported, or the baselines under-tuned.

• Maintain double-blind style in self-citations, platform names, acknowledgements, funding, and

the repository description.

Claim-discipline for recommender papers

• State the split protocol in words, not just in a config: "we use a temporal leave-one-last

split" pre-empts the leakage objection.

• Say the baselines were tuned with the same budget as the method; this one sentence defuses

the field's central reproducibility complaint.

• When an offline metric is the only evidence, scope the claim to offline; do not let "nDCG rose"

masquerade as "users are better served."

• Report variance (mean ± sd over seeds), and in captions say whether bars are sd, confidence

intervals, or quantiles.

• Label beyond-accuracy goals (diversity, fairness, exposure) as measured quantities, not

asserted virtues.

Sentence-level rewrites

| Draft pattern | RecSys-safe rewrite |

|---|---|

| "Our model significantly outperforms all baselines." | "improves nDCG@20 by X (sd Y) over equal-budget-tuned baselines" |

| "We evaluate on a standard split." | "We use a temporal leave-one-last split to avoid future leakage." |

| "Achieves state-of-the-art recommendation." | Claim scoped to the datasets and cutoff actually tested |

| "Users will benefit from better recommendations." | "our off-policy estimate of engagement rises; deployment is future work" |

| "We use Recall@20." | "Recall@20 over the full item catalog (not a sampled candidate set)" |

Vignette: compressing into eight two-column pages

A draft with a model, five datasets, and a sprawling related-work section: keep the method, the

assumption, the headline tuned table, one mechanism ablation, and the offline-online bridge in the

body; compress related work into contribution contrasts; move the tuning grid, two datasets, and

per-dataset breakdowns to the appendix with explicit forward references. The test of a good cut: a

reviewer should reconstruct the whole argument, including how the baselines were tuned, without

opening the repository.

Output format

[Writing diagnosis] clear / under-justified / overclaimed / leakage-ambiguous
[First-page fix] <new recommendation-first framing>
[Claim discipline] <claim -> table / ablation / off-policy result / scoped limitation>
[Evaluation wording] <split protocol / metric type / tuning-budget statement>
[Compression cuts] <move / delete / merge>
[Anonymity edits] <phrases to rewrite>