From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at Pinterest

Summary

Pinterest built a dedicated shopping candidate generation model to optimize for lower-funnel purchase conversions rather than relying strictly on engagement retrieval models. Because offsite conversion events are sparse and delayed, engineers trained a single multi-surface two-tower model using conversion data supplemented by duration-weighted click engagement and unengaged ad impressions as hard negatives. The two-tower retrieval architecture incorporates a parallel DCN v2 and MLP layer structure that allows simultaneous learning from the same inputs without information bottlenecks. Later revisions transitioned the system from a multi-head loss structure to a unified single-head multi-task architecture paired with an advertiser-level objective. Deployed across more than 600 million monthly active users, the system increased shopping conversion volume by 2.3% and boosted return on ad spend by 3.1% for United States shopping campaigns.

Context

Pinterest's shopping ads retrieval historically relied on engagement-based models that were not designed to optimize for lower-funnel conversions. Modeling offsite conversion actions like checkouts or add-to-carts is technically challenging because offsite advertiser-reported signals are significantly sparser, noisier, and more delayed than onsite user engagement data.

Approach / What changed

Pinterest built a multi-surface two-tower candidate generation model trained across Homefeed, Related Pins, and Search. The system supplements conversion data with click-duration-weighted engagement signals and unengaged ad impressions for negative sampling. Structurally, the towers employ parallel DCN v2 and MLP layers to eliminate sequential information bottlenecks, alongside a transition from a multi-head design to a unified single-head multi-task architecture featuring an advertiser-level loss objective.

Takeaways

• Replacing sequential DCN v2 and MLP cross layers with a parallel design yielded an 11% increase in offline recall@1000 by letting both networks learn directly from input features.
• Transitioning to a unified single-head multi-task architecture with an advertiser-level loss function delivered a 42% average increase in recall@100 over the 2023 baseline model.
• Deploying the conversion candidate generation system drove a 2.3% lift in shopping conversion volume, a 1.5% increase in CTR, and a 3.1% improvement in US shopping campaign RoAS.

Related reading

• Pinterest · Jul 27, 2026

  Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at Pinterest

  Pinterest designed the Pinner Progression initiative to shift its home feed recommendation pipeline from short-term engagement optimization toward long-term user retention. Standard ranking systems model user actions as immediate sequential features, failing to capture the evolving lifecycles of distinct user habits. To address this gap, the platform introduced User Interest Clusters (UICs), which cluster a user's engaged Pins in OmniSage embedding space using dynamic cluster counts and stateful lifecycle metadata. Pinterest integrated a UIC-aware penalty directly into Sliding Spectrum Diversity (SSD) scoring by matching Pin embeddings to cluster medoids via cosine similarity. Online experiments demonstrated that balancing use-case representation increased interaction diversity, extended session lengths, and delivered meaningful engagement gains across the platform.

  Pinterest Engineering

• Meta · Aug 3, 2026