From Clicks to Conversions: Architecting Shopping Conversion Candidate Generation at Pinterest (7 minute read)
Pinterest built a machine learning system that optimizes shopping ads for actual purchases rather than clicks, addressing the misalignment between engagement metrics and buying intent.
What: A two-tower neural network retrieval model that generates shopping ad candidates based on offsite conversion signals (actual purchases) instead of traditional engagement metrics. The system uses a multi-task architecture with DCN v2 and MLP cross layers, along with specialized training techniques to handle sparse conversion data.
Why it matters: Click-based optimization generates abundant training data but correlates poorly with whether users actually buy products, creating a fundamental problem for e-commerce advertising effectiveness.
Decoder
- Two-tower model: Neural network architecture with separate encoders for users and items that can be computed independently for efficient retrieval
- DCN v2: Deep & Cross Network version 2, a neural architecture designed to learn feature interactions
- Offsite conversions: Purchase events that happen on advertiser websites after clicking an ad, rather than on-platform engagement
Original article
Pinterest built a dedicated two-tower retrieval model to generate better shopping ad candidates optimized for offsite conversions, moving beyond traditional click/engagement-based signals which are abundant but poorly correlated with actual buying intent. The system uses a unified multi-task architecture with parallel DCN v2 and MLP cross layers, clever training techniques to handle sparse and noisy conversion data, and an advertiser-level loss function.