Much of Geospatial Web Search Is Beyond Traditional GIS

Official repository for the paper accepted as a full paper to COSIT 2026

We apply a SetFit binary classifier and UMAP+HDBSCAN clustering to the full MS MARCO corpus of 1.01M Bing queries, identify 181,827 geospatial queries (18.0%), and derive a taxonomy of 88 categories grouped into 9 themes.

Repository structure

Path	Contents
0-ProcessData.ipynb	Sample 5,000 queries via k-means++, weak-label with Llama 3.1, build gold dataset
1-TrainClassifier.ipynb	Train and evaluate the SetFit classifier; run on full MS MARCO
2-Cluster.ipynb	UMAP + HDBSCAN grid search, consistency checks, final clustering
queries.tar.gz	Raw MS MARCO queries (train + validation + test). This needs to be unarchived
gold-dataset/	1,200 hand-labelled queries, split into train/val/test as 200/200/800
interim/	Intermediate outputs (embeddings, cluster summaries)
output/	Final classifier predictions
visual/	Radial taxonomy chart source and other generated visuals

Reproducing the pipeline

Run the notebooks in order: 0-ProcessData → 1-TrainClassifier → 2-Cluster. Each notebook documents its own dependencies.

The classifier can be used directly without retraining:

from setfit import SetFitModel
model = SetFitModel.from_pretrained('ilyankou/is-geospatial-query')
model.predict(['restaurants near hyde park', 'what is greek yoghurt'])
# [1, 0]

Citation

@misc{ilyankou2026geospatialwebsearchtraditional,
      title={Much of Geospatial Web Search Is Beyond Traditional GIS}, 
      author={Ilya Ilyankou and Stefano Cavazzi and James Haworth},
      year={2026},
      eprint={2605.11336},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2605.11336}
}

Licence

Code released under the MIT licence. MS MARCO queries, including the gold dataset, are subject to the original MS MARCO terms.