Problem
An ads platform optimizing for conversions (checkouts, signups, leads) instead of clicks faces three obstacles. Labels come from advertiser sites, so volumes are abnormal and user matching and attribution are probabilistic. Conversions are far rarer than clicks, which limits model complexity and makes A/B tests need more traffic or time. They also arrive late after the click or view, so recent events look like false negatives in training.
Example from the article: Pinterest started in 2018 with a GBDT plus logistic regression hybrid. It then spent several years moving to multi-task deep networks, ensembles and user sequences, and finally to GPU serving.
Solution in brief
Pinterest trained one multi-task network on clicks and conversions, then added stronger feature interaction modules, an in-model ensemble, a shared bottom to cut cost, and a DHEN ensemble with a user-sequence transformer. Heavy models run on GPUs with CUDA graphs and half precision.
Questions this case answers
- How to raise the accuracy of a sparse-label, delayed-feedback prediction model when labels are noisy and the model must stay cheap to serve
- How to share signal from frequent upstream events with a rare downstream target in one model
- How to improve a sparse, delayed and noisy ads conversion prediction model step by step?
- How to combine click and conversion objectives in one multi-task ranking model?
Tools and methods
- Multi-task learning (MTL) — one model for clicks, good clicks, checkout and add-to-cart; frequent onsite actions help the sparse conversion tasks.
- DCNv2, Transformer, MaskNet — feature interaction modules that replace a plain MLP.
- In-model ensemble and DHEN — several feature interaction backbones combined inside one model.
- User sequence transformer — encodes long user activity histories.
- CUDA graphs and mixed precision (FP32/FP16) — GPU serving at lower latency and cost.
Google or a one-time email code. No password.