AI Friend Suggestions Just Got Smarter: 16% More Connections

When a social network suggests friends, it's not just looking at your profile. It looks at the whole web of who you know, who they know, and who they know. That's a massive puzzle when you have hundreds of millions of users and billions of connections. The AI behind this is called a graph neural network (a model that thinks in terms of connections), and it's normally very memory-hungry.

This paper, from researchers working on a production social graph with 194 million users and 28 billion friendship links, tackles two big problems. First: storing a unique "shorthand" (called an embedding) for each user used to require over 200 gigabytes of memory. They replaced that with a multi-hash trick — splitting each user's ID into a few compact codes — which slashed the memory by more than 98% while keeping recommendation quality high.

Second: when looking for who you might know, the AI used to scan someone's entire list of friends, which is impossible if you have tens of thousands of connections. Instead, they built a smart index that only checks recent interactions, using a fast search method (binary search, if you're into that) to find the most relevant friends in milliseconds. This is like checking just the last few months of your message history instead of reading your entire inbox.

In an online test, the new system increased friend additions from recommendations by 16% and the number of people actually adding friends by 11.5%. For you, that means the suggested people you see are more likely to be people you genuinely want to connect with — and it costs the company less to deliver them, which could mean faster, cheaper features for everyone.