Enhancing Ad Relevance: Integrating Real-Time Context into Sequential Recommender Models

This paper presents a Contextual Sequential Two-Tower Model for Pinterest ads that integrates real-time context into sequential recommender systems.

• •The model architecture integrates a context layer into the query tower that concatenates Transformer encoder outputs with context features like subject Pin embeddings and user demographics
• •Synthetic augmented data derived from positive labels is used during offline training to enable the model to learn from context information unavailable at training time
• •Hybrid inference approach splits computation: offline inference stores transformer outputs in feature store with daily refresh, while online inference computes context layer and final MLP at serving time
• •Offline evaluation demonstrates 3x to 10x improvement in Recall@K compared to production model, with median candidate relevance increasing by 275-300%