Place embeddings · Step 3
How IRIS learns the embedding
Nobody sat down and decided that dimension three should mean "busy high street". The eight numbers are learned, and the way they are learned is by comparison. IRIS reads millions of places at the same time and nudges their vectors so that places with similar surroundings move closer together and places that are different move apart. Do that across the whole continent, over and over, and the space organises itself: high streets drift toward high streets, quiet residential edges toward quiet residential edges, without anyone naming a single category.
Learned by comparison
This is the same kind of contrastive training that sits behind modern search and retrieval systems, where the model learns what "similar" means by being shown pairs that should be close and pairs that should be far apart. IRIS applies that idea to geography. A place and its immediate surroundings form a "should be close" signal; a randomly chosen distant place forms a "should be far" one. The encoder learns whatever eight-number code best satisfies millions of those comparisons at once.
Places with similar surroundings are pulled together. Different places are pushed apart. The eight numbers are whatever code satisfies millions of those comparisons at once.
Comparison alone can drift, so a few supervised checks keep the space anchored to things that matter: whether a certain kind of activity happens nearby, what sort of place this is. Those checks stop the embedding from organising itself around something real but useless, and keep it pointed at the differences a location decision actually turns on.
Similar places sit close, so types cluster on their own.
Each dot is a place. IRIS turns everything known about that place into a vector, and this view projects those vectors down to two dimensions. Because position follows similarity, places gather into soft neighbourhoods by type. Colour marks the place type. A city centre in one country lands beside a city centre in another, since similarity, not the national border, decides where a place sits.
A 2D shadow of the 8-dimensional space. Illustrative.
Encoder in, vector out, nothing rebuilt
It is worth being precise about the shape of the model, because an earlier telling of this story was wrong and we corrected it. IRIS uses an encoder built for tabular data: more than 100 signals go in, the encoder reads them, and eight numbers come out. There is no decoder. Nothing tries to reconstruct the original signals from the vector. The vector is the output, used directly. That is a simpler and more honest picture than the "compress then rebuild" story it replaced.
| Reconstruction: compress, then rebuild | IRIS: encoder-only |
|---|---|
| Compresses the input, then tries to rebuild the original | Reads the signals once; the eight-number vector is the output, used directly |
| Needs a decoder and a reconstruction step | No decoder, nothing rebuilt |
| Judged on how closely it reproduces its own input | Judged on putting genuinely similar places close together |
What the training does not give you
Here is the limit, stated plainly, because it is the part most likely to be oversold. The embedding is learned across places, not scored against a held-out set of stores it has never seen. That means it has no measured accuracy, no coverage figure, and no prediction interval of its own. Those are properties of a forecast tested against reality, and the embedding is not a forecast.
So we do not attach a percentage to it, and you should be wary of anyone who does. When IRIS quotes a range and a track record, that lives on the forecasting model downstream, checked on real openings, not on the embedding underneath. The embedding earns its trust a different way: by being inspectable, by putting genuinely similar places close together in a way you can eyeball, and by being honest about the fact that its eight dimensions have no names.
Learned by comparison, anchored by a few checks, and clear about what it has not been tested for. The next question is a practical one: what grid does IRIS lay all of this on, and why a hexagon.