Ads Ranking Evolution at Pinterest - From Logistic Regression to Deep Learning

Explore the evolution of ads ranking at Pinterest in this 53-minute podcast episode featuring Aayush Mudgal, Senior Machine Learning Engineer. Gain insights into the transition from traditional logistic regressions to deep learning-based transformer models, incorporating sequential signals, multi-task learning, and transfer learning. Discover the challenges faced and lessons learned in scaling ads ranking using innovative machine learning algorithms and platform advancements. Learn about the importance of continuous innovation, data transformation, monitoring, and pipeline optimization in large-scale recommendation systems. Understand the complexities of migrating to new technologies, optimizing model complexity, and balancing performance with cost considerations in the context of Pinterest's ads marketplace.

[] Join the AI in Production Conference!
[] Aayush preferred coffee
[] Takeaways
[] Evolving decision-making based on evolution and ROI
[] Different companies have varying approaches to building
[] Continuous innovation and adaptation are key. Embrace change
[] Transform, train, and analyze data for effective predictions
[] Shift in traditional systems, monitoring, and visibility
[] Monitoring all pipelines, models, features, and predictions. Alerting
[] Maintain simplicity and optimize pipelines for scaling
[] Check if systems are ready for change
[] Commitment, tooling, and understanding are crucial for migration
[] Concerns about technology support and migration strategy
[] Difficulty removing hybrid systems, but speed benefit.
[] Recommendation models learn user-content interactions, transformer as a feature interaction layer
[] Optimize model complexity, control sequence length, and reduce costs
[] Pinterest uses Pytorch for training and complex serving
[] Wrap up