Researchers Just Made 'People You May Know' Suggestions Smarter
Better 'people you may know' — and sharper fraud detection too. Here's how.
Think of any network: your friends on a social app, companies trading with each other, proteins interacting in your body, or accounts moving money around. A central question in network science is "link prediction" — guessing which two nodes (a person, a company, a protein) are likely to form a connection that isn't visible yet. That guess is exactly what powers "people you may know" suggestions, fraud-ring detection, and drug research. Two researchers, Yingying Zhang and Chengye Zhao, have published a new way to make those guesses better.
The old standby is "local similarity": if you and I share five friends, we're probably connected. It's cheap and fast, but it fails in sparse networks — places where most people barely know anyone. There, two nodes with no mutual friends both score zero, and the method shrugs. The fix is to add "node centrality," a measure of how important someone is across the whole network. PageRank, the idea behind early Google search, is the famous example. Importance plus shared friends gives a fuller picture.
Earlier work fused PageRank with just one similarity measure. This paper generalizes it to seven, then tests the combinations on nine real networks. It also introduces a newer centrality measure called DomiRank, creating a family of blended scores labelled DR-MD. Crucially, DomiRank doesn't need each measure tuned by hand. Across all nine datasets, the DomiRank blends beat both the plain baselines and the PageRank versions. One variant, DR-RA, reached an average accuracy score (AUC) of 0.7084 — where 0.5 is a coin flip and 1.0 is perfect.
The catch: this is an academic preprint, not a product. No app, no code release, and the improvement is real but modest — 0.7084 is good, not magic. It was tested on existing datasets, not live platforms with billions of users. Still, the framework is built to be reused with other measures, so expect smarter recommendations, sharper fraud detection, and faster drug-discovery hunches to trickle out over the next few years.
- Link prediction is the math behind 'people you may know,' fraud-ring detection, and drug research — guessing which two things will connect next
- The new method mixes two scores: who's important network-wide, and who already shares friends; it beat older approaches on all nine test networks
- It's a preprint, so no product or app yet, and the gain is real but modest — 0.7084 versus a 0.5 coin flip
Why It Matters
Better friend suggestions and fraud detection could arrive in a few years — not next week.