Unifying Ads Engagement Modeling Across Pinterest Surfaces

Summary

The article presents a comprehensive approach to unify ads engagement modeling across different surfaces at Pinterest, addressing the challenges posed by previously independent models. It outlines the fragmentation in user sequence modeling, feature representation, and training configurations that led to inefficiencies. The authors describe their strategy for creating a unified engagement framework, emphasizing the importance of maintaining operational safety, iterative development, and surface-specific calibration. Key architectural changes include the integration of multi-task learning and surface-specific exports, which enhance flexibility and performance while reducing operational costs. The article concludes with a discussion on future milestones aimed at further unifying the engagement model.

Key Learnings

• 1Unifying models can significantly reduce operational inefficiencies and improve iteration velocity by consolidating similar functionalities.
• 2Surface-specific calibration is crucial for accurate predictions, as different surfaces may have distinct traffic distributions.
• 3Implementing a multi-task learning design allows for flexibility in adapting surface-specific features while benefiting from shared learning.
• 4Efficiency improvements, such as request-level broadcasting and simplified compute paths, are essential to manage infrastructure costs effectively.
• 5Strategic architectural changes can lead to better performance metrics without compromising on cost-effectiveness.

Who Should Read This

Senior Machine Learning Engineers focusing on model optimization and scalability in ad tech environments.

Test Your Knowledge

Topics