Evolution of Ads Conversion Optimization Models at Pinterest

-Label quality: The quality of conversion labels is often dependent on advertisers, which can lead to inaccurate or abnormal conversion volumes. This limits the complexity of the models that can be built.

-Evolution to deep learning: The evolution of Pinterest's conversion optimization models began with a gradient boosting decision tree (GBDT) model combined with logistic regression. This was then transitioned to a deep learning-based single model, allowing for multi-task learning (MTL) by training multiple objectives together.

-Feature interaction modules: Modern recommendation systems require effective feature interaction learning modules. Pinterest implemented a low-dimensional cross network and deep network to capture low-ordered and implicit feature interactions, respectively.

-Multi-task Learning and Ensemble Learning: MTL was introduced to combine multiple conversion objectives into a unified model, leveraging onsite actions such as clicks to enhance training. In-model ensemble techniques were also implemented, combining the DCNv2 and transformer model backbones for feature crossing.

-Decoupling feature interaction modules: To reduce serving infrastructure cost, feature interaction modules were decoupled from feature processing modules and a shared bottom architecture was utilized.