Wanted to share my experience with a small experiment:
I trained a TabularModel on historic real-estate sales data available from the Israeli IRS specifically in center Tel Aviv.
Then I saved the embedding of the street address, and found out that the model learned the “map” of Tel-Aviv, without me providing any coordinates or map related data.
Features were: sale date, sale amount, apartment size, building year it was built…
Attached image shows selected street (Lincoln st) and it’s nearest points which on map are streets which intersect with the selected street.