recsys-topic-selection

RecSys Topic Selection

Use this before writing. RecSys is a single-domain venue: it is strongest when the central

claim is about recommendation — ranking objectives, user/item modeling, offline/online

evaluation, feedback loops, exposure and fairness, or deployed recommender behavior. Being merely

applicable to recommendation is not enough; the contribution has to speak to a recommender

audience.

Fit test

• Prefer RecSys when the main claim is a recommendation result: a ranking objective, a

user/session model, an evaluation protocol, an off-policy/counterfactual method, a

fairness/diversity/exposure result, or a deployed-system insight.

• Route to SIGIR when the core is ad-hoc retrieval, search ranking, or IR evaluation not tied

to recommendation.

• Route to KDD when the contribution is a general data-mining or large-scale algorithm and

recommendation is just one application.

• Route to WSDM / TheWebConf (WWW) when the emphasis is web-scale search-and-mining or web

systems broadly.

• Route to UAI or an ML venue when the contribution is a general learning/probabilistic

advance rather than recommender-specific evidence.

• Route to CHI / CSCW when user or community outcomes dominate over the recommendation

algorithm.

Fit signal table

| Signal in the project | RecSys reading |

|---|---|

| A ranking/user-modeling idea evaluated with tuned baselines and a leakage-aware split | Core fit — the house genre |

| Off-policy or counterfactual evaluation of recommendations | Core fit — a RecSys distinctive |

| A deployed system with production constraints and A/B evidence | Core fit — route to the Industry track |

| A reproduction/refutation of prior recommender results | Core fit — route to the Reproducibility track |

| A general retrieval or mining method, recommendation as one demo | Better at SIGIR / KDD / WSDM |

| A general ML method with a recommender benchmark tacked on | Better at an ML venue |

Which RecSys track

• Main long paper: a rounded recommendation contribution with offline (and ideally online)

evidence.

• Main short / Past-Present-Future: one focused finding, or a reflective/forward-looking position.
• Reproducibility: repeating, refuting, or re-scoping prior results; a dataset or framework.
• Industry: a deployed system with production constraints and live evidence.
• Resource / Dataset: a community dataset or software resource with build methodology.

Vignette: where an exposure-correction method goes

A project delivers an off-policy ranker with an exposure-corrected estimator and a simulator

bridge. RecSys reading: strong fit — a recommendation-specific evaluation advance is exactly what

the venue rewards, Main long paper. Strip the recommendation framing and keep only a generic

off-policy estimator, and it drifts toward an ML venue; turn it into a reproduction of three

published rankers, and the Reproducibility track becomes the right home.

Sharpening moves before committing

• Name the recommendation primitive: the ranking objective, the user/item model, the evaluation

protocol, or the deployment claim. If none exists, the RecSys framing does not.

• Confirm the evidence can meet the venue's evaluation bar (tuned baselines, leakage-aware split,

reported variance) — decoration-only benchmarks are a quiet fit failure here.

• Topic emphasis and the track lineup drift between cycles (2026 dropped LBR, added R&P Notes);

scan the current CFP before final routing.

Output format

[Fit] strong RecSys / possible RecSys / better elsewhere
[Best venue] RecSys / SIGIR / KDD / WSDM / TheWebConf / UAI / CHI / ML venue / other
[RecSys track] main-long / main-short / past-present-future / reproducibility / industry / resource
[Contribution sentence] <one sentence>
[Top rejection risk] <novelty / evaluation validity / scope / fit>
[Next action] <experiment, framing, or venue switch>