RecSys Related Work
Use this to audit novelty and eligibility. Reopen the current Call for Contributions for
dual-submission, anonymity, and prior-publication rules before advising authors.
Positioning checks
- Separate the recommendation contribution from generic ML improvement: a new ranking
objective, user/item modeling idea, evaluation protocol, off-policy estimator, or empirical
insight about deployed behavior.
- Cover the recommendation subfields your claim touches: collaborative filtering, sequential/
session models, off-policy/counterfactual evaluation, fairness/diversity/exposure, and — for any
empirical claim — the reproducibility-critique line.
- Treat ACM Digital Library, journal, and formal conference proceedings as archival unless current
rules say otherwise.
- Cite arXiv and prior-version work so double-blind review survives; do not point reviewers to
identity-revealing pages.
- Explain overlap with any concurrent or prior version, and do not submit duplicate archival work.
Neighbor-venue coverage table
RecSys is a single-domain venue, but its neighbors publish recommendation-relevant work. A
reviewer checks whether you cite the right neighbor, not just RecSys itself.
| Neighbor venue |
What it contributes to your related work |
Reviewer check |
| SIGIR |
Ranking models, IR evaluation, neural retrieval |
Did you distinguish recommendation from ad-hoc retrieval framing? |
| KDD |
Large-scale mining, the sampled-metrics critique |
Is the scalability or evaluation-methodology neighbor acknowledged? |
| WSDM / TheWebConf |
Web-scale recommendation, graph and CF methods |
Is the nearest web-recommendation method compared? |
| UAI |
Foundational ranking (e.g., BPR) and probabilistic modeling |
Are canonical recommendation methods cited to their true venue, not RecSys? |
| Reproducibility line |
"Are we really making much progress?" and follow-ups |
Does your evaluation answer the tuned-baseline critique head on? |
A bibliography that cites only RecSys papers tells a reviewer you may have missed the method that
already solved this at a neighbor venue — a recognizable reject pattern that benchmark strength
does not repair.
Positioning vignette
Imagine the paper proposes an exposure-corrected session ranker. Its nearest neighbors: a SIGIR
neural ranker with no exposure correction, a KDD paper on sampled-metric bias, and a prior RecSys
counterfactual-embedding paper. The novelty sentence should name all three contrasts — exposure
handling where the SIGIR line ignored it, full-ranking evaluation answering the KDD critique, and
a session-level objective where the prior RecSys work was static.
Concurrent-work judgment calls
- Independently concurrent arXiv work: cite neutrally, state the technical difference, avoid
priority claims reviewers cannot verify.
- Your own prior version: verify its archival status against the current CFP and phrase the
citation so double-blind review survives.
- When unsure whether a venue is archival, declare the overlap on the submission form rather than
gambling on a chair's interpretation.
Output format
[Eligibility] clear / needs declaration / risky
[Closest subfields] <CF / sequential / off-policy / fairness / reproducibility>
[Nearest 3 works] <work -> distinction (with true venue)>
[Archival-overlap risk] <none / issues>
[Novelty sentence] <RecSys-ready recommendation contribution contrast>
1---2name: recsys-related-work3description: Use when positioning an ACM RecSys submission against recommender-systems literature and its neighbors (SIGIR, KDD, WSDM, TheWebConf, UAI), including the reproducibility-critique line, arXiv and prior versions, concurrent work, ACM Digital Library archival status, and the citation coverage RecSys reviewers expect across recommendation subfields.4---56# RecSys Related Work78Use this to audit novelty and eligibility. Reopen the current Call for Contributions for9dual-submission, anonymity, and prior-publication rules before advising authors.1011## Positioning checks1213- Separate the **recommendation** contribution from generic ML improvement: a new ranking14 objective, user/item modeling idea, evaluation protocol, off-policy estimator, or empirical15 insight about deployed behavior.16- Cover the recommendation subfields your claim touches: collaborative filtering, sequential/17 session models, off-policy/counterfactual evaluation, fairness/diversity/exposure, and — for any18 empirical claim — the **reproducibility-critique** line.19- Treat ACM Digital Library, journal, and formal conference proceedings as archival unless current20 rules say otherwise.21- Cite arXiv and prior-version work so double-blind review survives; do not point reviewers to22 identity-revealing pages.23- Explain overlap with any concurrent or prior version, and do not submit duplicate archival work.2425## Neighbor-venue coverage table2627RecSys is a single-domain venue, but its neighbors publish recommendation-relevant work. A28reviewer checks whether you cite the *right* neighbor, not just RecSys itself.2930| Neighbor venue | What it contributes to your related work | Reviewer check |31|---|---|---|32| SIGIR | Ranking models, IR evaluation, neural retrieval | Did you distinguish recommendation from ad-hoc retrieval framing? |33| KDD | Large-scale mining, the sampled-metrics critique | Is the scalability or evaluation-methodology neighbor acknowledged? |34| WSDM / TheWebConf | Web-scale recommendation, graph and CF methods | Is the nearest web-recommendation method compared? |35| UAI | Foundational ranking (e.g., BPR) and probabilistic modeling | Are canonical recommendation methods cited to their true venue, not RecSys? |36| Reproducibility line | "Are we really making much progress?" and follow-ups | Does your evaluation answer the tuned-baseline critique head on? |3738A bibliography that cites only RecSys papers tells a reviewer you may have missed the method that39already solved this at a neighbor venue — a recognizable reject pattern that benchmark strength40does not repair.4142## Positioning vignette4344Imagine the paper proposes an exposure-corrected session ranker. Its nearest neighbors: a SIGIR45neural ranker with no exposure correction, a KDD paper on sampled-metric bias, and a prior RecSys46counterfactual-embedding paper. The novelty sentence should name all three contrasts — exposure47handling where the SIGIR line ignored it, full-ranking evaluation answering the KDD critique, and48a session-level objective where the prior RecSys work was static.4950## Concurrent-work judgment calls5152- Independently concurrent arXiv work: cite neutrally, state the technical difference, avoid53 priority claims reviewers cannot verify.54- Your own prior version: verify its archival status against the current CFP and phrase the55 citation so double-blind review survives.56- When unsure whether a venue is archival, declare the overlap on the submission form rather than57 gambling on a chair's interpretation.5859## Output format6061```text62[Eligibility] clear / needs declaration / risky63[Closest subfields] <CF / sequential / off-policy / fairness / reproducibility>64[Nearest 3 works] <work -> distinction (with true venue)>65[Archival-overlap risk] <none / issues>66[Novelty sentence] <RecSys-ready recommendation contribution contrast>67```
brycewang-stanford