TechPrufer
Get started
Pinterest
Pinterest logoPinterest · Feature trail

Shopping Conversion Candidate Generation

2 posts · Apr 2026 – Sep 2026

What this trail covers

Pinterest evolved its candidate generation for shopping conversion ads by introducing a 3-tower engagement co-train model. This model jointly predicts pCTR, pGCTR30, and pOCTR, leveraging a two-stage scoring flow that combines fast two-tower predictions with richer three-tower predictions for a selected subset of candidates. Key architectural components include a cross tower for user sequence and candidate feature interaction, and an inter module for shared representation learning. Serving optimizations, such as moving Pin pre-processing into the Pin tower and using lower precision, reduced inference latency significantly. The new model achieved substantial offline loss reduction for three-tower predictions and translated to online lifts in CTR and gCTR30, while maintaining two-tower prediction quality and reducing cost per click.

All Pinterest trails
  1. Beyond Two Towers: Launching the 3-Tower Engagement Co-Train Model (Part 2)This post details the development and implementation of a 3-tower engagement co-train model for shopping conversion ads, building upon the existing Nexus serving stack. The model architecture introduces a cross tower for richer interaction modeling and an inter module for shared representation, co-trained with a fast two-tower prediction head. Key technical contributions include partitioning query and Pin embeddings for task-specific segments, a weighted multi-task loss function, and significant serving optimizations like moving Pin pre-processing into the Pin tower and exploring lower precision inference. The model achieved ~30% reduction in offline loss for three-tower predictions and maintained comparable quality to the production two-tower model for fast scoring.
  2. From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at PinterestThis post details the architectural evolution of Pinterest's shopping conversion candidate generation model. Key contributions include the design of a parallel DCN v2 and MLP cross layers architecture, which improved offline recall by +11% and was adopted across production engagement retrieval models. It also describes the transition from a multi-head to a unified single-head multi-task architecture, incorporating an advertiser-level loss function to better stabilize query embeddings in low conversion coverage regions. This led to an average increase of +42% recall@100 for conversion tasks compared to the previous model.