Pinterest unified three separate ads engagement models for Home Feed, Search, and Related Pins into one shared architecture.
•Three diverged surface models caused low iteration velocity, redundant training costs, and high maintenance overhead
•Unification followed three principles: start simple with merged components, iterate incrementally with surface-aware features, and maintain operational safety
•Surface-specific calibration layers for HF and SR outperformed a single global calibration layer
•Multi-task learning with surface-specific checkpoint exports enabled shared representation while allowing per-surface iteration
•Efficiency improved via DCNv2 projection layers, fused kernel embedding, TF32 training, and request-level user embedding broadcasting
Pinterest Engineering Blog - Medium