Geo Embeddings: Universal Representations for Locations

Olga Tikhobaeva, Elena Shulyak25 May 2024

Problem

Geospatial models in the bank each use their own set of tables keyed to map squares. Every new task needs a fresh collection of features and a fresh EDA, and there is no common baseline. Prototyping is slow, and raw per-square data cannot be handed to partners in joint projects.

Example from the article: the bank's map has 207,875 squares of 500×500 m. Each square carries client counts, app entries, average check and real estate data. The team wants one anonymized vector per square that stands in for all of them.

Solution in brief

Train a vector for every grid square from two sources. One is the bank's tabular square features, turned into a knowledge graph and embedded by link prediction. The other is OpenStreetMap layers drawn as multi-channel images and embedded by loc2vec or a masked autoencoder. The vectors are then used as features, alone or next to ordinary ones.

Questions this case answers
  • How to build one reusable vector per spatial cell from tabular cell statistics and map layers, so that downstream models need no per-task feature tables
  • How to embed the numeric values of a table with a link-prediction graph embedding model so that the raw values need not be shared
  • How to build reusable location embeddings from tabular and map data for downstream models
  • How to replace dozens of per-grid-cell feature tables with one compact vector

Tools and methods

  • Relational data embeddings (KEN) — table to knowledge graph to embeddings by link prediction with negative sampling
  • loc2vec — triplet-loss network on 14-channel map images, ResNet50
  • ViT MAE — masked autoencoder on the same images
  • OpenStreetMap, Geofabrik extracts — source of buildings, roads, rail, water and public transport layers
Continue reading — it's free

Google or a one-time email code. No password.