Skip to content

SYSTEM Cited by 1 source

Pinterest Sequential Two-Tower CG

Definition

Pinterest's Sequential Two-Tower CG is the predecessor candidate generator to systems/pinterest-contextual-sequential-cg — a two-tower retrieval model using a Transformer-based encoder over the user's offsite conversion history to predict future interactions with advertisers and specific products. It was "a significant step forward, moving beyond static interest categories to model the evolving user shopping journey" (sources/2026-05-08-pinterest-enhancing-ad-relevance-integrating-real-time-context-into-sequential-recommender-models).

The original system is described in a prior Pinterest post — Ads Candidate Generation using Behavioral Sequence Modeling — that is not separately ingested on this wiki as of 2026-05-21. This page is bounded to what the contextual-evolution post discloses about the baseline.

Architectural shape (as disclosed in the contextual evolution post)

  • Two-tower retrieval model.
  • User tower: Transformer encoder over a sequence of offsite conversion events (purchases / checkouts on advertiser sites). The Transformer's output is the user embedding.
  • User embedding inferred offline only — "The user embeddings were inferred offline purely from historical offsite behavior, meaning that at the moment an ad was served, the model had no knowledge of what the user was currently browsing on Pinterest."
  • Candidate generation for ads — used to surface advertisers and specific products the user is likely to engage with.

Known limitation that motivated the contextual evolution

On Pinterest's contextual surfaces (Related Pins and Search) — where the "current Pin" or search query is a strong real-time intent signal — purely offline user embeddings could not compete:

"In the previous production system, less than 1% of impressions on Related Pins were attributed to this CG, indicating its candidates struggled to survive the downstream ranking and auction stages."

This is a survival-rate failure: the CG retrieved candidates that the downstream ranking + auction kept dropping because they didn't reflect the immediate session intent. Contextual Sequential CG is the structural fix.

Relationship to other Pinterest ads CGs

  • systems/pinterest-contextual-sequential-cg — direct successor. Same Transformer-over-offsite-history substrate; adds context layer, synthetic pseudo-context training, hybrid offline/online inference.
  • systems/pinterest-shopping-conversion-cg — sibling CG with different lineage (shopping ads, parallel DCNv2+MLP cross layers, multi-task with engagement auxiliary, advertiser-level loss). Both target offsite conversions but with different architectural primitives.

Caveats

  • Stub — The original 2024-era Ads Candidate Generation using Behavioral Sequence Modeling post is not ingested on the wiki. This page captures only what the 2026-05-08 contextual-evolution post discloses about the baseline.
  • Topology undisclosed — Transformer layer count, sequence length, embedding dimension, training-loss specifics not described in the contextual-evolution post.
  • Surface deployment — Used on Related Pins (where the survival-rate problem manifested); deployment on other surfaces (Home Feed, Search) not documented in this post.

Seen in

Last updated · 542 distilled / 1,571 read