Biased graph sampling for better related-product recommendation
230%. That's how much Amazon improved AI recommendations by tweaking one graph sampling technique.

Why it matters
Amazon's research shows that biased graph sampling can dramatically improve AI recommendation systems, offering a concrete path for companies to enhance their machine learning performance with existing data.
The key facts
5 to know230% improvement in graph-neural-network embeddings utility
Biased graph sampling technique
Tailored neighborhood sizes and sampling probability
Related-product recommendation optimization
Amazon Science research
Go to the source
Amazon Scienceamazon.science
Publisher excerpt: Tailoring neighborhood sizes and sampling probability to nodes’ degree of connectivity improves the utility of graph-neural-network embeddings by as much as 230%.


