SASRec
The self-attentive sequential recommender that replaced RNNs — still the baseline every new recsys paper compares against.
About
SASRec (Self-Attentive Sequential Recommendation) is the 2018 ICDM paper that ended the RNN era in sequential recommendation. Wang-Cheng Kang and Julian McAuley showed that a transformer with causal masking — attend only to past actions — captures long-term user dynamics like an RNN while making predictions from relatively few relevant actions like a Markov chain, and trains an order of magnitude faster than comparable CNN/RNN models thanks to parallel self-attention.
At each step the model identifies which items in the user's history are actually relevant and uses them to predict the next one, and the attention-weight visualizations turned out to be genuinely interpretable. The authors released TensorFlow code, the community ported it to PyTorch, and it has been the default baseline in sequential-recommendation papers ever since — TIGER, BERT4Rec follow-ups, and dozens more all report their SASRec numbers.
It is a research reference, not a product: there is no hosted API, and on its own it does not model time gaps or item content, which is exactly what TiSASRec, UniSRec, and RecFormer were built to add.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (ICDM 2018): outperformed state-of-the-art MC/CNN/RNN sequential models on both sparse and dense datasets; reported an order-of-magnitude training efficiency gain over comparable RNN/CNN models
- No independent benchmark body re-scores it — its standing comes from thousands of follow-up papers reproducing it as the baseline
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (TIER verified, metrics-v1). Evidence: peer-reviewed ICDM 2018 paper by Kang and McAuley with author-released TensorFlow code (github.com/kang205/SASRec) and mature PyTorch ports; reproduced as the standard baseline across thousands of subsequent papers over 8 years. Strengths: foundational architecture, fully open code, unmatched replication record, still practically useful as a lightweight baseline. Docked for: (1) author-reported numbers only, and it is old enough that newer models routinely beat it; (2) maintenance is frozen — the value is the recipe, not active development. License: open research code.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 10.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · ICDM 2018 paper (Kang & McAuley) with author-released code at github.com/kang205/SASRec and mature PyTorch ports; the standard sequential-recsys baseline for 8 years.
- · Paper-reported SOTA over MC/CNN/RNN models on sparse and dense datasets with order-of-magnitude training efficiency gains.
- · Author-reported numbers only; maintenance frozen — the value is the recipe. Newer models routinely beat it.