From Clicks to Conversions: Designing Pinterest's Shopping Conversion Candidate Generation

Pinterest built a dedicated candidate generation model to shift shopping retrieval from click-centric to conversion-centric. The first model in 2023 lifted both conversion and engagement, and the 2025 improved version went on to achieve a 3.1% improvement in US shopping campaign RoAS.

The data design focused on compensating for sparse and delayed offsite conversion signals. Homefeed, Related Pins, and Search were trained together as a single multi-surface model, and a dual positive signals strategy was applied, using click and repin alongside conversion. Clicks had their noise reduced through log-based re-weighting that reflected dwell time, and ad impressions without engagement were used as harder hard negatives.

The model is built on a two-tower structure that separates the user tower and Pin tower, with DCN v2 layered on top, and placed DCN v2 and MLP in parallel instead of sequentially to reduce information bottlenecks. This change lifted offline recall@1000 by +11%, and was subsequently rolled out to production engagement retrieval models as well.

The multi-task architecture also changed.

• In 2023, a multi-head structure with separate conversion and engagement heads was used.
• In 2025, this was unified into a unified single-head multi-task architecture, so that the serving embedding reflects both objectives together.
• To compensate for the variance in Pin-level conversion signals, an advertiser-level loss was added, resulting in a +42% improvement in recall@100 for the conversion task compared to the 2023 model.

Ultimately, this candidate generation model led to a 2.3% increase in shopping conversion volume, a 2.7% rise in shopping impression-to-conversion rate, a 1.5% increase in CTR, and a 2.2% increase in CTR for sessions over 30 seconds. Pinterest demonstrated that conversion optimization and improving the Pinner experience can be advanced together without conflict.