Pinterest Unifies Ads Engagement Modeling Across Multiple Surfaces

Pinterest's ads engagement models were scattered across multiple surfaces like Home Feed (HF), Search (SR), and Related Pins (RP), but the core challenge was ultimately the same: predicting the same ads engagement as accurately as possible. Previously, independent production models were operated per surface, and differences in user sequence modeling, feature crossing, feature representation, and training configuration led to slower iteration speed, duplicated training costs, and growing maintenance burden.

To address this, a baseline unified model was built first. Features from the three surfaces were combined, existing modules were merged into a single architecture, and training data was also unified. Offline performance improved, but training and serving costs increased significantly, requiring further optimization for actual production deployment.

In the next step, RP, which had relatively higher costs, was excluded, and HF and SR were unified first. In this process, key components like MMoE and long user sequences were combined into a single unified model, and elements that showed unstable effects when applied to a single surface alone led to better improvements when trained jointly on HF+SR. The final target architecture is one model supporting all three surfaces, but performing only the computation appropriate to each surface through a surface-specific tower tree and its modules.

In terms of accuracy, surface-specific calibration was important. A single global calibration layer could mix together the traffic distributions of HF and SR, so this was changed to a view type specific calibration layer that calibrates each separately, and online experiments showed better performance than the existing shared calibration.

Multi-task learning and surface-specific exports were also introduced to restore flexibility. While maintaining a single shared architecture, surface-specific checkpoints were exported separately so that each surface could choose the structure more suitable for it, while still retaining the benefits of shared representation learning.

Infrastructure optimizations were also carried out in parallel to reduce performance and cost issues.

• DCNv2 was used to project the Transformer output into a smaller representation before connecting it to downstream crossing and the tower tree, lowering serving latency.
• fused kernel embedding improved inference latency, and TF32 boosted training speed.
• On the serving side, request-level broadcasting was applied so that embedding lookups for the same user were not repeated—fetched once and then broadcast across the entire batch.

Finally, improved performance for HF and SR was confirmed in both offline experiments and online experiments. The next step is to extend the model to RP, which has not yet been unified, and the biggest challenge will be meeting serving performance targets while maintaining higher efficiency.