
Innovating Ad Performance and User Experience through Parallel DCN v2 and Multi-Task Learning
This article details Pinterest's journey in architecting a specialized candidate generation model for shopping conversions. It explores how they evolved from engagement-based retrieval to a sophisticated system capable of capturing high-intent purchase signals among 600 million users.
Highly recommended for ML engineers focusing on conversion optimization in large-scale recommendation systems and architects seeking to improve retrieval quality.
Pinterest faced challenges in optimizing for lower-funnel conversions because offsite conversion signals are sparse, noisy, and delayed compared to onsite engagement signals.
The team implemented a two-tower model with parallel DCN v2 and MLP cross-layers, utilizing weighted engagement data based on click duration and an advertiser-level loss function.
The system achieved a 2.3% increase in shopping conversion volume, a 3.1% improvement in RoAS, and an average +42% increase in recall@100 for conversion tasks.
Trade-off
Addressing data sparsity required more complex multi-task balancing and advertiser-level granularity, which increased architectural complexity compared to engagement-only models.
A model architecture that combines a cross network for explicit feature interactions and a deep network for implicit patterns.
A retrieval architecture where user and item features are encoded separately into embeddings for efficient similarity matching.
A learning paradigm that trains a model on multiple related tasks simultaneously to improve generalization and handle data sparsity.




