A place embedding is a dense, fixed-length vector representation of a geographic location that captures its semantic character — the kinds of activities that happen there, its relationship to surrounding places, and its position within the urban fabric. Place embeddings extend the representational power of word embeddings (like Word2Vec and BERT) to geographic space, treating locations as entities that can be learned from their spatial and contextual co-occurrences.
From Word Embeddings to Place Embeddings
The intuition behind word embeddings — that words with similar contextual distributions should have similar representations — translates naturally to space. The distributional hypothesis in NLP (“you shall know a word by the company it keeps”) becomes a spatial distributional hypothesis: you shall know a place by the company it surrounds itself with.
A café that is consistently near offices, transit stations, and other cafés develops a representational fingerprint different from one surrounded by parks and playgrounds — even if both are categorised identically in a database. Place embeddings capture this distinction.
How Place Embeddings Are Learned
Co-occurrence methods treat GPS traces or mobility sequences as “sentences” of locations. A model like Place2Vec applies Skip-Gram training: given a visit to location v, predict the nearby locations within a spatial or temporal window. Locations that appear in similar contexts receive similar embeddings.
Graph-based methods build a spatial graph (H3 cell graph, road network, or POI proximity graph) and apply graph neural network encoders. Each node’s embedding is iteratively updated by aggregating information from its neighbours — analogous to how BERT aggregates context from surrounding tokens.
Multi-modal fusion combines multiple data sources: OpenStreetMap features, satellite imagery patches, street-view images, and social media activity. Each modality contributes a partial signal about a place’s character; a fusion layer learns to weight and combine them.
What Gets Encoded?
Well-trained place embeddings encode a surprisingly rich set of geographic properties:
- Functional use — residential, commercial, industrial, recreational, mixed
- Urban density — the built environment intensity of the surrounding area
- Accessibility — proximity to transit, walkability, cycling infrastructure
- Temporal rhythm — which hours of day are active (learned from mobility data)
- Socioeconomic character — correlates with deprivation indices in held-out tests
Crucially, these properties are learned from unlabelled data — the model is never told what a place’s function is; it infers it from the spatial distribution of OSM features and mobility patterns.
Evaluation Challenges
Evaluating place embeddings is harder than evaluating word embeddings. The canonical NLP analogy task (“king − man + woman = queen”) doesn’t translate cleanly to geography. Standard evaluation tasks include:
- Land-use classification — do embeddings cluster by functional type?
- Crime/deprivation prediction — do embeddings predict held-out socioeconomic indicators?
- Transfer to new cities — do embeddings trained on London generalise to Berlin?
The last criterion is particularly important for real-world utility and remains an open research problem.
Applications
Urban planning — place embeddings reveal functional zones without manual digitisation. Planners can identify areas undergoing functional change (gentrification, industrial decline) before it appears in official land-use registers.
Real estate — embedding similarity can find analogous neighbourhoods in different cities, supporting comparative investment analysis.
Humanitarian response — embedding models trained on high-income cities have been applied (with caution) to estimate service accessibility in data-sparse contexts.
Personalised routing — integrating a user’s mobility history with place embeddings enables routing that prefers the kinds of places a user typically visits.
MORPHEME
My MORPHEME project learns place embeddings at the H3 cell level for 24 global cities using a self-supervised GNN trained on OpenStreetMap feature distributions. The resulting embeddings encode urban character — vibrancy, safety, and function — without relying on any labelled training data, enabling direct comparison of neighbourhoods across cities and continents.
Related work: WalkGrid uses a similar feature-embedding approach to match walking environments to user preferences across 150,000 H3 cells. See the full SAWiP research programme for how place embeddings are applied to conflict and displacement data.