Skip to content

CONCEPT Cited by 1 source

Shopping conversion candidate generation

Definition

Shopping conversion candidate generation is the retrieval-stage ads-ML primitive that narrows a large shopping-Pin (or shopping-product) catalog to a rank-tractable set of high-purchase-intent candidates — optimised for offsite conversions (checkout, add-to-cart) rather than onsite engagement (click, repin, save).

It is a decomposition point in the ads funnel: the retriever feeds retrieval → L1 ranking → L2 ranking → auction, and separating conversion retrieval from engagement retrieval lets the retriever's loss function + positives + negatives + architecture be tuned for a structurally different objective than engagement.

Why a dedicated retriever

Engagement-based shopping retrieval pipelines maximise onsite interaction. That's the wrong optimisation target for conversion-optimised campaigns (oCPM for conversions, RoAS-focused advertisers) because:

  1. Conversion signals are sparse, noisy, advertiser-reported, and delayed — treating conversion as a subcase of engagement under-weights it and lets the denser engagement signal dominate gradient direction.
  2. Candidate pools for conversion and engagement can genuinely differ. High-conversion-intent Pins aren't the same set as high-engagement Pins — a Pin that gets many saves/clicks may not drive purchases.
  3. Advertiser performance metrics (RoAS) depend on conversion-accuracy retrieval ceilings. Retriever recall is the upper bound on what the ranker can score; a conversion-blind retriever caps advertiser outcomes.

Design levers specific to conversion CG

Different from a generic engagement retriever, a conversion CG typically needs:

  • Dual positives. Supplement primary conversion positives with engagement positives to broaden coverage. Pattern: dual positive signal.
  • Auxiliary-task regularisation. Engagement prediction as auxiliary task stabilises the shared trunk via abundant gradient signal. Pattern: auxiliary engagement task.
  • Coarser-granularity loss. Per-Pin conversion data is high-variance; an advertiser-level loss adds a parallel training objective at a more stable granularity.
  • Hard negatives. In-batch negatives alone don't cover the boundary of served-but-not-converting ads — adding served-ad-no-engagement hard negatives helps the model learn the decision boundary of the inventory actually shown.
  • Multi-surface single model. Sparse labels can't be fragmented across surface-specific models; a shared multi-surface model with surface-specific features concentrates label budget.

Canonical wiki instance — Pinterest (2023-2025)

Pinterest's Shopping Conversion Candidate Generation system (sources/2026-04-27-pinterest-from-clicks-to-conversions-architecting-shopping-conversion-candidate-generation) is the wiki's canonical instance:

  • Separate retriever from the engagement CG pipeline — conversion-centric loss + positives + negatives.
  • Two-tower retrieval architecture, parallel DCNv2 + MLP cross layers in both towers.
  • Multi-task heads with engagement as auxiliary task for training stability.
  • 2025 evolution to unified single-head multi-task with advertiser-level loss → +42% recall@100 for conversion tasks, +3.1% RoAS on US shopping campaigns.

Production wins (Pinterest US, 2023-2025): +2.3% shopping conversion volume, +2.7% shopping impression-to-conversion rate, +1.5% CTR, +2.2% CTR over 30s, +3.1% RoAS.

Relationship to adjacent concepts

Caveats

  • Parallel vs sibling retrieval pipelines. Pinterest runs the conversion CG alongside an engagement-based shopping retrieval pipeline; the post doesn't describe how their outputs merge in the downstream funnel.
  • Not generic advertising retrieval — specific to shopping-intent ads where conversion (checkout / add-to-cart) is the advertiser's target, not brand-awareness or top-of-funnel impression buys.

Seen in

Last updated · 445 distilled / 1,275 read