UniSRec
One pretrained sequence model for many domains — learns universal item representations from text instead of IDs.
About
UniSRec attacks the most annoying property of classic recommenders: an ID-based model trained on one catalog is useless on another. Published at KDD 2022 by Yupeng Hou and colleagues at Renmin University's AI Box, it represents every item by its text description, encoded with BERT and then adapted through a lightweight parametric-whitening plus mixture-of-experts module that smooths out the very different vocabularies of different domains ("sweet, fresh" for food vs "war, story" for movies).
Two contrastive pretraining tasks — sequence-to-item and sequence-to-sequence with multi-domain negatives — produce a universal sequence representation that transfers to new domains and even new platforms with parameter-efficient fine-tuning, in both inductive and transductive settings. The paper reported strong cross-domain performance without heavy retraining, and the authors released full PyTorch code and preprocessing scripts.
It is the direct ancestor of the text-based wave that followed (RecFormer, and conceptually TIGER's semantic IDs), and the pragmatic choice when you need one model across multiple catalogs.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (KDD 2022, arXiv 2206.05941): strong cross-domain and cross-platform transfer performance on real-world datasets with parameter-efficient fine-tuning, outperforming ID-based transfer baselines
- No independent benchmark body re-scores it; subsequent papers (e.g. SISA-Rec) use it as the text-based transfer baseline
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: peer-reviewed KDD 2022 paper (arXiv 2206.05941) from Renmin University with maintained PyTorch code, preprocessing scripts, and mid-product data releases at github.com/RUCAIBox/UniSRec; cited as the baseline in later transfer-learning recsys work. Strengths: open code plus useful released artifacts, genuinely novel cross-domain framing, influential on later text-based models. Docked for: (1) author-reported numbers only; (2) semantic encoding is kept separate from the attention mechanism, which later work flags as a limitation. License: open research code.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · KDD 2022 paper (arXiv 2206.05941) from Renmin University RUC AI Box: text-based universal sequence representations via BERT + parametric whitening + MoE adaptor, two contrastive pretraining tasks.
- · Maintained repo with preprocessing scripts and mid-product data releases; cited as baseline in later transfer-learning recsys work.
- · Author-reported numbers only; semantic encoding kept separate from attention, a noted limitation.