Tools for geocoding the provenance of web information.
Python JavaScript HTML CSS
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


This project contains tools for inferring the geoprovenance of webresources. A web page's geoprovenance is the home country for the original publisher of the information contained in the web page.

The geoprovenance of:

  • A web resource created by a company or organization is the country where its headquarters are located.
  • A web resource created by an individual author is that individual's home country.
  • A book is the country associated with the publisher of the book.
  • More details are provided in our 2015 CHI paper.

If you use this software in an academic publication, please cite it as follows: Sen, S., Ford, H., Musicant, D., Graham, M., Keyes, O., and Hecht, B. 2015. "Barriers to the Localness of Volunteered Geographic Information." Proceedings of CHI 2015. New York: ACM Press.

There are a few deviations from the published algorithm (TODO fix these):

  • Does not follow links to and other repositories.
  • Does not handle binary files, like PDFs, Open Office docs, and Microsoft docs.
  • Not integrated with amazon web services to scrape en-mass.

Installing necessary Python modules:

pip install tldextract
pip install chardet
pip install geopy
pip install beautifulsoup4
pip install langid

If you want to run the evaluator, which rebuilds the logistic regression (not necessary to use the pre-built model), you'll also need to install sklearn.

Running the command-line program.

The py directory contains the program, which reads URLs from standard input and writes information about them to standard output. For example, if you ran the following from the shell from within the py directory:

$ echo '' | python ./

You would see the following output:  gb      0.8259  {'gb' : 0.8259, 'us' : 0.0628, 'fr' : 0.0008, 'ca' : 0.0006, 'ru' : 0.0006, 'in' : 0.0005, 'de' : 0.0005, 'se' : 0.0005, 'it' : 0.0005, 'pl' : 0.0005} outputs the following four tab-separated fields:

  1. The URL itself.
  2. The most probable country.
  3. The estimated probability the most probable country is correct (in this case, about 83%).
  4. The top 10 candidate countries, along with their associated probabilities, in JSON format.

If you run the program from somewhere outside of the py directory, or would like to use a larger pre-built feature cache (see information below), you can specify the feature directory, or both the features and data directories, from the command line:

$ python path/to/features/dir
$ python path/to/features/dir path/to/data/dir

Incorporating the module into your own Python program.

import gputils
import gpinfer

# necessary iff not run from the "py" directory

# necessary iff not run from the "py" directory or alternate feature caches are used (see below)

inferrer = gpinfer.LogisticInferrer()

# conf is a number between 0 and 1.0 indicating confidence
# dist is a dict with keys country codes and values predicted probability
(conf, dist) = inferrer.infer('')

Incorporating larger pre-built caches for speed

A larger feature cache is available at To use this feature cache, download and extract the zip file. You'll then need to point the module at the feature directory by either specifying the appropriate argument to the program, or by calling gp_utils.set_feature_dir with the appropriate absolute pathname.

This cache contains information about all 7.5M URLs analyzed in our CHI paper.

The GeoProv198 Dataset

The logistic regression classification model used in this package is trained using a gold standard dataset that maps urls to countries. This dataset is available in the data directory and its collection methodology is described in the citation above.

Questions or suggestions?

Open an issue on this repo, send a pull request, or email Shilad at


  • Shilad Sen developed the geoprovenance inference algorithm and software.
  • Dave Musicant developed the original Ruby code to extract country names from whois queries.
  • Heather Ford led development of GeoProv198, with major assistance from Brent Hecht and minor assistance from Dave Musicant, Shilad Sen, and Mark Graham.
  • Matthew Zook provided early guidance on the design of our algorithm.