This AI Predicts Where People Go — Even in Cities It's Never Seen

When AI tries to predict how people move around a city — where you'll head next, which roads will clog at 8am — it first needs to turn real places into something a computer can compare. That's called a "location embedding": a list of numbers that stands in for a spot on the map, the way a phone number stands in for a person. The problem is that most existing versions only work if you already have lots of data about how people actually travelled there.

LE4Mob, a new research method from a team of UK and Chinese academics, flips that. It learns what a place is like from geography and map context rather than from tracing people's movements. Three things make it different. It's "inductive," meaning it can make sense of a brand-new neighbourhood it has never seen data for. It's "distance-aware," meaning it keeps real-world distances intact — two places an hour apart stay an hour apart in the maths. And it's "general-purpose," so the same representation can be reused for different jobs instead of being rebuilt each time.

The team tested it on two very practical tasks: guessing an individual's next stop, and generating commuter flows between areas at the population level. Across several datasets and cities, LE4Mob beat the standard alternatives — and its lead was biggest when the data was thin, or when locations were new and unseen. That matters, because thin data is the normal situation for small towns and fast-changing neighbourhoods.

So what's the catch? This is a 13-page academic preprint, not a product. It needs good underlying map and geographic data to work, and it hasn't been turned into anything you can download. But the direction is significant: the smarter version of urban prediction here runs on public maps rather than on surveillance of your personal movements — meaning better traffic tools and planning without quite so much watching.