/doppelganger

A Python package of tools to support population synthesizers

Primary LanguagePythonApache License 2.0Apache-2.0

Doppelganger

A Python package of tools to support population synthesizers. Population synthesizers create make-believe or synthetic households and persons for use in agent-based models, i.e. models or simulations that attempt to represent the behavior of individual actors or "agents".

Merriam Webster's fourth definition of "synthetic": devised, arranged, or fabricated for special situations to imitate or replace usual realities

Test status Coverage Status

Installing

Doppelganger relies on numpy. You can check if you have numpy on your system by running this from your shell:

pip freeze | grep numpy

If you see a result like numpy==[version_number], you're good to go.

If you don't already have numpy running on your system, we strongly recommend you use a version of python with numpy pre-built. We recommend Anaconda but there are several options.

Installing for use

To use simply use the doppelganger package, run

pip install doppelganger

You should now have doppelganger installed.

Installing for development

To develop Doppelganger, once you have numpy set up, clone or download this repository, go to the root directory and run

pip install .[tests]

You should now have the local doppelganger code & testing tools installed.

From the repo root, run the tests to confirm everything is working:

py.test

Examples

The example directory contains a Jupyter Notebook of examples and demonstrations of Doppelganger's features.

Features

Doppelganger has the following two key feature categories that we hope will improve population synthesis in practice:

  • Bayesian Networks. A Bayesian network is a directed graph of conditional probabilities for a set of random variables. Doppelganger allows users to easily build Bayesian nets from the Public Use Microdata Sample (PUMS) data that is collected and distributed by the Census Bureau. Once created, these Bayesian nets can be traversed to create synthetic households and persons that have the same relationships, in the aggregate, as the source (in this case the PUMS) households and persons. Bayesian nets have the ability to both (a) add heterogeniety from synthetic populations drawn from a small sample of observations and (b) obscure the specific attributes of the observed sample.
  • Convex Optimization. Doppelganger, like most population synthesizer packages, allows users to allocate a set of observed or synthetic (created with Bayesian nets) households or persons to a geography such that the aggregate characteristics of the synthetic set matches the aggregate characteristics the user believes to be true about the geography. For example, it allows the user to allocate individual households to a PUMA such that the income distribution of the collection of households matches the income distribution from another data source (e.g., other summaries of the American Community Survey or Decennial Census). Doppelganger uses convex optimization to solve the allocation problem. Convex optimization has the attractive features of (a) generating a consistent solution, when one is available, and (b) allowing the user to introduce subjective weights to either prioritize one set of controls over another and/or efficiently overcome inconsistent controls.

What's Next?

Doppelganger version 0.1 is the beginning of our work -- a simple demonstration of the potential uses of the key features -- with population synthesizers, not the end. Up next: adding features and improving performance.

Interested in Collaborating?

We'd love to hear from and collaborate with you.

  • If you're a government agency interested in deploying a population synthesizer, we'd be interested in understanding your needs and workflow to guide our next development push.
  • If you're an academic and/or model developer, we'd love to hear what enhancements would make your research or next deployment more successful.
  • If you're a developer, please share anything cool you're doing or would like to do with the toolkit.

Please communicate with us via GitHub's Issues.

Credits

Doppelganger is inspired by:

Licensing

Apache 2.0 © Sidewalk Labs