Strengthening Ad Relevance: Integrating Real-Time Context into Sequential Recommendation Models

The existing ad candidate generator (CG) was a Transformer-based two-tower model that only used off-site conversion history, failing to reflect the context the user is currently viewing. In particular, on surfaces like Related Pins and Search where current context matters most, candidates failed to pass ranking and auction, and on Related Pins this CG's share of impressions was under 1%.

To address this, Pinterest introduced a Contextual Sequential Two Tower Model, adding a context layer to the query tower. Real-time context is combined with the sequence representation produced by the Transformer encoder to form the final user embedding, and on Related Pins, the interest categories of the subject Pin (the Pin currently being viewed) weighted by confidence score were used as the context input. Demographic embeddings such as age, country, and gender were also added to the user representation.

Since actual online context couldn't be used directly for training, training was conducted with synthetic augmented data. Pseudo-context was generated from positive labels (conversion events) and injected into the input, and context layer dropout was increased during training so that sequence information continued to be learned as well.

Serving was structured as hybrid inference.

• Offline: The last hidden state of the Transformer encoder was stored in the feature store and refreshed daily.
• Online: Only the context layer and the final MLP head were computed at ad request time.

The results were clear. Offline Recall@K was 3x to 10x higher than the production model, and as the survival rate of candidates passing through the subsequent ranking funnel increased, median relevance improved by about 275-300%. The overall ad relevance metric for Related Pins rose 1.08%, and the number of candidates that reached impressions also doubled. Ultimately, ROAS rose by about 0.7%, and in top countries with a large share of revenue it rose by 1.4%.

Going forward, Pinterest plans to extend this contextual CG to Search, and is considering combining context and sequence more sophisticatedly via cross-attention instead of simple concatenation.