Pinterest Boosts Ad Performance Using Graph Neural Networks and Off-Platform Conversion Data

Researchers at Pinterest have developed a machine learning framework that links users' activity on the platform with their shopping behavior off-site to improve ad targeting and performance. By constructing a large-scale graph from both onsite interactions and opted-in offsite conversions, the team created a more complete picture of user purchasing intent.

Graph Neural Networks Meet Real-World Ads#

Graph-based learning systems have become standard in recommendation engines across the tech industry, but most focus solely on activity within a single platform. Pinterest's approach extends this model by explicitly incorporating how users actually spend money after leaving the site. The framework builds a heterogeneous graph connecting ads, users, and conversion signals across boundaries.

A New Embedding Approach#

The team introduced TransRA (TransR with Anchors), a specialized knowledge graph embedding model designed to translate graph patterns into numerical representations that ads ranking systems can use directly. Early attempts to plug these embeddings into the ranking pipeline yielded disappointing results, so the researchers refined the method using two key innovations: a large ID embedding table and an attention-based fine-tuning approach that learns how to weight the graph signal during training.

Measurable Business Impact#

When deployed in Pinterest's Ads Engagement Model, the system delivered concrete improvements: a 2.69% lift in click-through rate and a 1.34% reduction in cost-per-click. These gains came from the model's ability to recognize user interests more accurately by synthesizing both what people do on Pinterest and what they actually purchase elsewhere.

The techniques are now part of Pinterest's production ad system and represent a wider lesson for other large-scale recommendation platforms seeking to improve predictions by bridging online and offline customer behavior.