Problem
Pinterest's shopping ads retrieval used engagement-based models, which were not built to optimize lower-funnel actions such as checkout or add-to-cart. Those conversions happen on the advertiser's site and are reported by the advertiser, so labels are far sparser, noisier and later than onsite clicks. A retrieval model trained on them alone sees too few positives, while one trained on clicks alone optimizes the wrong thing.
Example from the article: the first shopping conversion candidate generation model, launched in 2023, raised shopping conversion volume by 2.3% and click-through rate (CTR) by 1.5%.
Solution in brief
Train one two-tower retrieval model on conversions, supplemented with duration-weighted clicks and repins as an auxiliary task, across all shopping surfaces. Add a parallel DCN v2 and MLP cross layer to both towers, and merge separate conversion and engagement heads into one head with an extra advertiser-level conversion loss.
Questions this case answers
- How to train a retrieval model on a rare, noisy and delayed target when a frequent but weaker signal is available as auxiliary supervision
- How to improve the interaction capacity of a two-tower retrieval model while keeping serving latency low
- How to train a candidate generation model for purchase intent when conversion labels are sparse, noisy and delayed?
- How to use abundant click data without drowning the rare conversion signal?
Tools and methods
- Two-tower model — user/query and Pin towers, dot-product retrieval.
- DCN v2 (Deep & Cross Network v2) in parallel with an MLP — richer feature interactions per tower.
- Multi-task learning — conversion plus engagement objectives.
- Sampled softmax with hard negatives — impressions without engagement as negatives.
- Click-duration weighting — down-weights short, likely accidental clicks.
Google or a one-time email code. No password.