Scaling Recommendation Systems with Request-Level Deduplication

Engineers at **Pinterest Engineering**, including Sr. Machine Learning Engineer **Matt Lawhon**, Machine Learning Engineer II **Filip Ryzner**, Machine Learning Engineer II **Kousik Rajesh**, Sr. Staff Machine Learning Engineer **Chen Yang**, and Principal Engineer **Saurabh Vishwas Joshi**, detailed their work titled **Scaling Recommendation Systems with Request-Level Deduplication**. The team highlighted their **Foundation Model**, which earned an **oral spotlight** at **ACM RecSys 2025**.

The work addresses the mechanics of scaling recommendation models to improve the quality of content served to users. At **Pinterest**, scaling recommendation models delivers an outsized impact on serving quality, leveraging **request-level deduplication** as a key mechanism within production recommendation systems.

For infrastructure and machine learning engineers, the direct relationship between model size and recommendation quality makes scaling a primary engineering objective. Implementing **request-level deduplication** allows engineering teams to scale large systems like the **Foundation Model** while serving high-quality content to users.

In large-scale content platforms, serving quality depends on the capability to continuously scale recommendation infrastructure. Securing an **oral spotlight** at **ACM RecSys 2025** highlights **Pinterest**'s focus on advancing recommendation architecture through **request-level deduplication**.

Engineers building large recommendation pipelines should evaluate **request-level deduplication** strategies alongside foundation model scaling. Technical teams should monitor the detailed findings and presentation of the **Foundation Model** at **ACM RecSys 2025** to evaluate implementation patterns for high-throughput recommendation systems.