The self-organizing map.
The language in a listing is rarely literal. A listing that calls a kitchen "chef's" and a listing that calls it "updated" are pointing at different markets, not different appliances. To catch that, the texts were turned into vectors with a sentence transformer, which places phrases close together in a high-dimensional space when they mean similar things, not when they share words. Those vectors were then arranged on a 15 by 15 grid so that similar texts ended up near each other. Each cell on the grid below is a neuron, labeled with the most distinctive term among the listings that landed in it.
Neighborhoods as dialects of a market.
The whole point of organizing listings this way is to track how neighborhoods change over time. A tract that sat in one part of the map in 2004 and a different part in 2022 has not just rebuilt its houses. It has started speaking a different market dialect. Below, the same 15 by 15 grid is grouped into a smaller number of market clusters.
Reading the grid
The color of each cell is the k cluster that cell was assigned to. Cells near each other have similar language being used to describe them.
Four neighborhoods, told in their own words.
These four tracts illustrate different kinds of movement through the clusters. The paths below are built from the listings, period by period.
What comes next.
This page is a preview of the method, not the findings. Everything above was built on a subset of listings to show that the pipeline works and that the map it produces is readable.
The full analysis will run on the complete MLS corpus. The paper coming out of this work, co-authored with Elizabeth Delmelle and Isabelle Nilsson, will trace neighborhood trajectories and what the shifts in market dialect say about how an area has changed. Look out for it.