---
license: fair-noncommercial-research-license
language:
- en
---

# A multi-view contrastive learning framework for spatial embeddings in risk modeling

In this repository, we provide the pretrained models as described in our paper: 

> Holvoet, F., Blier-Wong, C., & Antonio, K. (2025). A multi-view contrastive learning framework for spatial embeddings in risk modeling. *arXiv preprint arXiv:2511.17954*.

Paper is available as pre-print via Arxiv: **[arXiv preprint arXiv:2511.17954](https://arxiv.org/abs/2511.17954)**

This repository accompanies our [GitHub repository](https://github.com/freekholvoet/MultiviewSpatialEmbeddings), where you can find the code to train the models and an example of usage.

## Using the pretrained models

The pretrained models described in Section 3.4 of the paper are provided in this repository.

There are five different models available:

- `EU16_GS32_OSM16.ckpt`
- `EU16_OSM16.ckpt`
- `EU32_GS96_OSM32.ckpt`
- `EU64_GS64.ckpt`
- `EU8_GS32_OSM32.ckpt`

Example on how to download the models and calculate embeddings for a list of latitude and longitude coordinates:

```python
from huggingface_hub import hf_hub_download
from load_lightweight import get_mvloc_encoder
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Example coordinates (latitude, longitude) of various European cities
c = torch.tensor([
    (50.8503, 4.3517),   # Brussels
    (48.8566, 2.3522),   # Paris
    (51.5074, -0.1278),  # London
    (52.5200, 13.4050),  # Berlin
    (41.9028, 12.4964),  # Rome
    (40.4168, -3.7038),  # Madrid
    (59.3293, 18.0686),  # Stockholm
    (60.1699, 24.9384),  # Helsinki
    (47.4979, 19.0402),  # Budapest
    (48.2082, 16.3738),  # Vienna
], dtype=torch.float32) 

model = get_mvloc_encoder(
    hf_hub_download("FreekH/multiview_spatial_embedding", "MODEL_NAME.ckpt"),
    device=device
)
model.to(device)

with torch.no_grad():
    emb = model(c.to(device).double()).detach().cpu().numpy()
```

Replace `MODEL_NAME.ckpt` with the desired model filename from the list above. The [GitHub repository](https://github.com/freekholvoet/MultiviewSpatialEmbeddings) contains a Jupyter Notebook, called Add_embeddings_to_data.ipynb, that includes a function to systematically add embeddings to a data set containing a latitude and a longitude feature. 

## Citation

Citing the paper:

```bibtex
@article{holvoet2025multiview,
  title={A multi-view contrastive learning framework for spatial embeddings in risk modeling},
  author={Holvoet, Freek and Blier-Wong, Christopher and Antonio, Katrien},
  journal={arXiv preprint arXiv:2511.17954},
  year={2025}
}
```

Citing the models: 

```bibtex
@misc{holvoet_pretrainedmodels,
	author       = { Freek Holvoet },
	title        = { Spatial embeddings via multiview contrastive learning},
	year         = 2025,
	note         = {[Pretrained spatial embedding models]},
	url          = { https://huggingface.co/FreekH/multiview_spatial_embedding },
	doi          = { 10.57967/hf/7009 },
	publisher    = { Hugging Face }
}
```