Most recommendation systems focus on ranking the main item—whether it’s the perfect product, ad, or video. But in high-impact environments like Meta, optimizing the presentation of that item can be just as critical. Which image best complements an ad? Which text overlay (“Best Seller,” “Limited Time”) will drive engagement? These “non-product” recommendations are often overlooked, yet they can be harder to model than the primary recommendation itself.
In this talk, I’ll share our team’s multi-year journey from early missteps to a robust, scalable ranking pipeline for creative elements that surround the main recommendation. We’ll explore:
Data collection and labeling strategies tailored to secondary recommendation targets
Model architectures and feature engineering approaches that work in large-scale environments
Lessons learned on experimentation speed, evaluation metrics, and failure patterns
Attendees will leave with actionable insights to design and deploy ranking models for any “artifact” that supports the main recommendation—whether in ads, e-commerce, or content platforms—so they can improve impact without overhauling their primary recommender system.
Supported by
