Place embeddings · Step 1
How a place becomes a vector
A language model turns a word into a vector: a list of numbers where words that behave alike, "king" and "queen", "Paris" and "Berlin", end up close together. IRIS does the same thing for a place. For one small patch of ground it produces a short list of numbers, and patches that behave alike end up close together in that number space. The analogy is exact where it counts, both are embeddings, and it stops there: a language model reads discrete words, while IRIS reads a location straight from more than 100 continuous signals about it. There are no words and no tokens in the middle. A place goes in, a vector comes out.
| A language model | An IRIS place embedding |
|---|---|
| Reads discrete words and tokens | Reads a location straight from more than 100 continuous signals |
| Turns a word into a vector | Turns one 120-metre patch of ground into a vector |
| Words that behave alike end up close together | Places that behave alike end up close together |
| The vector's dimensions have no plain-English names | Its eight dimensions have no plain-English names either |
What the vector actually holds
Each IRIS embedding is eight numbers. That is unusually few, and a specialist will raise an eyebrow, but it is the real figure: the model was trained to pack the signal that matters into a compact code rather than a long one. Those eight numbers are not eight scores you could read off a dashboard. They are learned dimensions, and on their own they have no names.
The dimensions have no names. Together they are the place's fingerprint, and two places with similar fingerprints behave alike.
This is the part it is tempting to overstate, so we will not. IRIS does not learn a "footfall number" in dimension one and a "wealth number" in dimension two. It learns whatever arrangement of eight numbers best separates places that behave differently and groups places that behave alike. If we describe a particular cell as "dense, mixed-use, busy by day", that is our reading of the pattern, offered so you can sanity-check it, not a label the model assigns.
Dozens of layers of a place, compressed into one.
Start with the ground itself: an aerial view of the streets, the blocks, the green space. Over that same footprint IRIS plots layer after layer of what is actually there, who lives and works nearby, how people move, what the roads carry, where the shops and stops are.
More than 100 signal layers stack up. Then IRIS compresses the whole pile into a single embedding: one hex layer that carries all of it at once, so two places can be compared on everything, not one column at a time.
- Aerial footprint
- Population and daytime density
- Footfall and traffic
- Points of interest
- Land use
- and 90+ more signal layers
- IRIS embedding
Read straight in, with nothing rebuilt
There is one more honest point about the mechanism. Some representation models work by trying to rebuild their input: compress it, then reconstruct the original and see how close you got. IRIS does not do that. It is encoder-only. Signals go in, the encoder reads them, and the eight-number vector comes out and is used directly. There is no decoder and no reconstruction step. That keeps the story simple and true: the vector is a compression of the place, used as is.
What the vector is not
A place embedding is a similarity map, not a forecast. It tells you which real locations a proposed site most resembles. It does not, by itself, say what that site will earn. Turning similarity into a revenue number is a separate job, done by a transparent, adjustable model that reads the trading history of the closest real matches. We keep those two steps apart on purpose, because it is the only honest way to describe them.
It also carries no accuracy figure of its own. The embedding is learned across millions of places at once; it is not scored against a held-out set of stores, so there is no coverage number or prediction interval attached to it. Those belong to the forecasting model downstream, tested on real openings. Anyone who quotes you an accuracy figure for the embedding itself is describing something that was never measured.
So: a place becomes a vector, eight learned numbers that put similar places close together. The next question is what goes into those numbers in the first place.