PinnerFormer
Pinterest's production transformer that learns user embeddings from action sequences to predict long-term engagement.
About
PinnerFormer is what happens when a sequential recommender has to survive at Pinterest scale — over 400 million monthly users and billions of pins. Published at KDD 2022 and deployed in production since fall 2021, it trains a transformer over a user's recent pin engagements and compresses the result into a single user embedding.
The key departure from the usual next-action models is the objective: a dense all-action loss that predicts the user's positive engagement over a 14-day window after the embedding is generated, rather than just the next click. That let Pinterest run it on cheap daily-batch infrastructure instead of realtime streaming, while closing most of the quality gap with realtime embeddings. Online A/B tests showed substantial improvements in user retention and engagement versus Pinterest's previous user representation.
There is no public code release — this is a production system described in a paper, not a downloadable model — but it shaped how the industry thinks about batch-friendly sequential user modeling, and successors like PinRec build directly on it.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (KDD 2022, arXiv 2205.04507): offline Recall@10 up to 0.2293 with a 6-layer, 768-dim transformer; Pinterest's own A/B tests showed substantial gains in user retention and engagement over the prior user representation
- Independent follow-up (PinRec paper, arXiv 2504.10507): PinnerFormer unordered Recall@10 of 0.461 (Homefeed) / 0.412 (Related Pins) / 0.257 (Search) as the baseline that PinRec improved on by 10–30%
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 KDD 2022 paper (arXiv 2205.04507, Pancha/Zhai/Leskovec/Rosenberg) with production deployment confirmed (fall 2021) and A/B validation on retention and engagement; used as the comparison baseline in later published work (PinRec). Strengths: real production system at massive scale, rigorous ablations, heavily cited, genuinely influential design. Docked for: (1) no public code or weights — replication is on you. License: proprietary (Pinterest internal).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · KDD 2022 paper (arXiv 2205.04507, Pancha/Zhai/Leskovec/Rosenberg); deployed in production at Pinterest since fall 2021 with A/B validation on retention and engagement.
- · Paper-reported offline Recall@10 up to 0.2293; PinRec paper later used it as baseline (0.461 Homefeed unordered Recall@10) and beat it by 10–30%.
- · Heavily cited and influential; no public code — replication is on the reader. Proprietary license.