Ads Candidate Generation using Behavioral Sequence Modeling

Summary

The article outlines Pinterest's innovative approach to enhancing ad candidate generation through behavioral sequence modeling. By leveraging a transformer-based model, the team predicts user interactions with advertisers based on historical behavior, significantly improving ad relevance. The architecture employs a two-tower model, integrating user event sequences and advertiser representations, and utilizes advanced training techniques such as in-batch negatives and sampled softmax loss with log-Q bias correction. The results demonstrate a marked increase in conversion rates and a reduction in cost per action, showcasing the effectiveness of this model in a real-world setting.

Key Learnings

• 1The use of transformer-based models allows for effective encoding of user behavior, leading to improved ad targeting.
• 2Incorporating in-batch negatives and log-Q bias correction helps mitigate popularity bias and enhances personalization.
• 3Evaluating model performance through Recall@K metrics provides insights into the effectiveness of ad retrieval systems.
• 4The architecture's dual approach for item-level interaction prediction facilitates deeper personalization in ad delivery.
• 5Real-time updates to user representations can significantly enhance the relevance of ad candidates.

Who Should Read This

Senior Machine Learning Engineers focusing on ad technology and personalization strategies

Test Your Knowledge

Topics