This repository contains the report submitted for my Master thesis at EPFL in March 2019. This research was supervised by Olivier Verscheure, executive director at the Swiss Data Science Center and Jean-Philippe Nantel, team director at the Computer Research Institute of Montreal.

This work presents a method to generate embeddings that capture the semantic characteristics of geographic locations and their surroundings. To this end, a triplet network is trained in a self-supervised manner and with unlabelled raster data to learn a similarity metric between locations and encode them in vectors of floating point numbers. Combining raster data generated from OpenStreetMap and automated features engineering allows the embeddings to capture complex semantic information hidden in a latent dimension. It is demonstrated that our model can generate meaningful embeddings for tens of thousands of locations all over the province of Québec. Finally, a typical use-case is presented where these embeddings are used as additional features in a properties price prediction problem.

Images speak a thousand words. Here is a visualization of embedded locations reduced in a 3-dimensional metric space.