Pinterest is refining its ad targeting by using behavioral sequence modeling to enhance how it generates ad candidates. The company recognizes that traditional methods based on demographic data often miss the mark in a fast-moving environment where user interests change rapidly. The new approach leverages historical user behavior to predict future interactions with advertisers and products, aiming to deliver ads that resonate more with individual users. The team initially focused on predicting which advertisers a user is likely to engage with next. They developed a transformer-based model that analyzes the sequence of products a user has viewed or purchased. This model generates a personalized list of advertisers, improving the relevance of ads shown to users. The process involves an offline workflow that predicts the top 100 advertisers for each user based on their behavior and integrates these predictions into the ad serving system. Building on this, Pinterest is now also predicting specific products users will engage with. This requires a deeper understanding of user intent and product attributes. The two-tower model architecture supports this by creating detailed representations of individual product Pins. Given the scale of Pinterest's catalog, with over 1 billion items, the model incorporates both in-batch negatives and a sample of 20 million Pins to enhance training. Performance is evaluated using hit rates at various levels, ensuring a balance between effective retrieval and diversity in recommendations, ultimately aiming for higher user satisfaction and engagement with ads.