Zest Race Predictor

Zest Race Predictor (ZRP) is an open-source machine learning algorithm that estimates the race/ethnicity of an individual using only their full name and home address as inputs. ZRP improves upon the most widely used racial and ethnic data estimation method, Bayesian Improved Surname Geocoding (BISG), developed by RAND Corporation in 2009.

ZRP was built using ML techniques such as gradient boosting and trained on voter data from the southeastern U.S. It was then validated on a national sample using adjusted tract-level American Community Survey (ACS) data. (Model training procedures are provided.)

Compared to BISG, ZRP correctly identified:

• 25% more African-Americans as African-American

• 35% fewer African-Americans as non-African American

• 60% fewer Whites as non-White

ZRP can be used to analyze racial equity and outcomes in critical spheres such as health care, financial services, criminal justice, or anywhere there’s a need to impute the race or ethnicity of a population dataset. (Usage examples are included.) The financial services industry, for example, has struggled for years to achieve more equitable outcomes amid charges of discrimination in lending practices.

Zest AI began developing ZRP in 2020 to improve the accuracy of our clients’ fair lending analyses by using more data and better math. We believe ZRP can greatly improve our understanding of the disparate impact and disparate treatment of protected-status borrowers. Armed with a better understanding of the disparities that exist in our financial system, we can highlight inequities and create a roadmap to improve equity in access to finance.

Notes

This is the preliminary version and implementation of the ZRP tool. We’re dedicated to continue improving both the algorithm and documentation and hope that government agencies, lenders, citizen data scientists and other interested parties will help us improve the model.

Install

Install requires an internet connection. We recommend installing zrp inside a python virtual environment. The package has been tested on 3.7.4, but should likely work with 3.7.X.

pip install zrp

After installing via pip, you need to download the lookup tables using the following command:

python -m zrp download

Note: Due to the size and number of lookup tables necesary for the zrp package, total installation requires 3 GB of available space.

Data

Training Data

The models available in this package were trained on voter registration data from the states of Florida , Georgia, and North Carolina. Summary statistics on these datasets and additional datasets used as validation can be found here .

Consult the following to download state voter registration data:

• North Carolina

• Florida

• Alabama

• South Carolina

• Georgia

• Louisiana

American Community Survey (ACS) Data:

The US Census Bureau details that, “the American Community Survey (ACS) is an ongoing survey that provides data every year – giving communities the current information they need to plan investments and services. The ACS covers a broad range of topics about social, economic, demographic, and housing characteristics of the U.S. population. The 5-year estimates from the ACS are “period” estimates that represent data collected over a period of time. The primary advantage of using multiyear estimates is the increased statistical reliability of the data for less populated areas and small population subgroups. The 5-year estimates are available for all geographies down to the block group level.” ( Bureau, US Census. “American Community Survey 5-Year Data (2009-2019).” Census.gov, 8 Dec. 2021, https://www.census.gov/data/developers/data-sets/acs-5year.html. )

ACS data is available in 1 or 5 year spans. The 5yr ACS data is the most comprehensive & is available at more granular levels than 1yr data. It is thus used in this work.

Usage and Examples

To get started using the ZRP, first ensure the download is complete (as described above) and xgboost == 1.0.2

Check out the guides in the examples folder. Clone the repo in order to obtain the example notebooks and data; this is not provided in the pip installable package. If you’re experiencing issues, first consult our common issues guide.

Next, we present the primary ways you’ll use ZRP.

ZRP Predictions

Summary of commands:

>>> from zrp import ZRP
>>> zest_race_predictor = ZRP()
>>> zest_race_predictor.fit()
>>> zrp_output = zest_race_predictor.transform(input_dataframe)

Breaking down key commands

>>> zest_race_predictor = ZRP()

• ZRP(pipe_path=None, support_files_path=”data/processed”, key=”ZEST_KEY”, first_name=”first_name”, middle_name=”middle_name”, last_name=”last_name”, house_number=”house_number”, street_address=”street_address”, city=”city”, state=”state”, zip_code=”zip_code”, race=’race’, proxy=”probs”, census_tract=None, street_address_2=None, name_prefix=None, name_suffix=None, na_values=None, file_path=None, geocode=True, bisg=True, readout=True, n_jobs=49, year=”2019”, span=”5”, runname=”test”)

  • What it does:

    • Prepares data to generate race & ethnicity proxies

  You can find parameter descriptions in the ZRP class and it’s parent class.

>>> zrp_output = zest_race_predictor.transform(input_dataframe)

• zest_race_predictor.transform(df)

  • What it does:

    • Processes input data and generates ZRP proxy predictions.

    • Attempts to predict on block group, then census tract, then zip code based on which level ACS data is found for. If Geo level data is unattainable, the BISG proxy is computed. No prediction returned if BISG cannot be computed either.

Parameters
	df : {DataFrame} Pandas dataframe containing input data (see below for necessary columns)

Input data, df, into the prediction/modeling pipeline MUST contain the following columns: first name, middle name, last name, house number, street address (street name), city, state, zip code, and zest key. Consult our common issues guide to ensure you’re input data is the correct format.

• Output: A dataframe with the following columns: AAPI AIAN BLACK HISPANIC WHITE source_block_group source_zip_code source_bisg

   >>> zrp_output
  
  =========== =========== =========== =========== =========== =========== ===================== ====================== ==================
               AAPI        AIAN        BLACK       HISPANIC    WHITE       source_block_group    source_census_tract    source_zip_code
  =========== =========== =========== =========== =========== =========== ===================== ====================== ==================
   ZEST_KEY
   10          0.021916    0.021960    0.004889    0.012153    0.939082    1.0                   0.0                    0.0
   100         0.009462    0.013033    0.003875    0.008469    0.965162    1.0                   0.0                    0.0
   103         0.107332    0.000674    0.000584    0.021980    0.869429    1.0                   0.0                    0.0
   106         0.177411    0.015208    0.003767    0.041668    0.761946    1.0                   0.0                    0.0
   109         0.000541    0.000416    0.000376    0.000932    0.997736    1.0                   0.0                    0.0
   ...         ...         ...         ...         ...         ...         ...                   ...                    ...
   556         NaN         NaN         NaN         NaN         NaN         0.0                   0.0                    0.0
   557         NaN         NaN         NaN         NaN         NaN         0.0                   0.0                    0.0
  =========== =========== =========== =========== =========== =========== ===================== ====================== ==================
  

One of the parameters to the parent class that ZRP() inherits from is file_path. This parameter allows you to specify where the artifacts/ folder is outputted during the run of the ZRP. Once the run is complete, the artifacts/ folder will contain the outputted race/ethnicity proxies and additional logs documenting the validity of input data. file_path need not be specified. If it is not defined, the artifacts/ folder will be placed in the same directory of the script running zrp. Subsequent runs will, however, overwrite the files in artifacts/; providing a unique directory path for file_path will avoid this.

ZRP Build

Summary of commands

>>> from zrp.modeling import ZRP_Build
>>> zest_race_predictor_builder = ZRP_Build('/path/to/desired/output/directory')
>>> zest_race_predictor_builder.fit()
>>> zrp_build_output = zest_race_predictor_builder.transform(input_training_data)

Breaking down key commands

>>> zest_race_predictor_builder = ZRP_Build('/path/to/desired/output/directory')

• ZRP_Build(file_path, zrp_model_name = ‘zrp_0’, zrp_model_source =’ct’)

  • What it does:

    • Prepares the class that builds the new custom ZRP model.

Parameters
	file_path : {str} The path where pipeline, model, and supporting data are saved.
	zrp_model_name : {str} Name of zrp_model.
	zrp_model_source : {str} Indicates the source of zrp_modeling data to use.

You can find more detailed parameter descriptions in the ZRP_Build class. ZRP_Build() also inherits initlizing parameters from its parent class.

>>> zrp_build_output = zest_race_predictor_builder.transform(input_training_data)

• zest_race_predictor_builder.transform(df)

  • What it does:

    • Builds a new custom ZRP model trained off of user input data when supplied with standard ZRP requirements including name, address, and race

    • Produces a custom model-pipeline. The pipeline, model, and supporting data are saved automatically to “~/data/experiments/model_source/data/” in the support files path defined.

    • The class assumes data is not broken into train and test sets, performs this split itself, and outputs predictions on the test set.

Parameters
	df : {DataFrame} Pandas dataframe containing input data (see below for necessary columns)

Input data, df, into this pipeline MUST contain the following columns: first name, middle name, last name, house number, street address (street name), city, state, zip code, zest key, and race. Consult our common issues guide to ensure you’re input data is the correct format.

• Output: A dictionary of race & ethnicity probablities and labels.

As mentioned in the ZRP Predict section above, once the run is complete, the artifacts/ folder will contain the outputted race/ethnicity proxies and additional logs documenting the validity of input data. Similarly, defining file_path need not be specified, but providing a unique directory path for file_path will avoid overwriting the artifacts/ folder. When running ZRP Build, however, artifacts/ also contains the processed test and train data, trained model, and pipeline.

Addition Runs of Your Custom Model

After having run ZRP_Build() you can re-use your custome model just like you run ours. All you must do is specify the path to the generated model and pipelines (this path is the same path as ‘/path/to/desired/output/directory’ that you defined previously when running ZRP_Build() in the example above; we call this ‘pipe_path’). Thus, you would run:

>>> from zrp import ZRP
>>> zest_race_predictor = ZRP('pipe_path')
>>> zest_race_predictor.fit()
>>> zrp_output = zest_race_predictor.transform(input_dataframe)

Validation

The models included in this package were trained on publicly-available voter registration data and validated multiple times: on hold out sets of voter registration data and on a national sample of PPP loan forgiveness data. The results were consistent across tests: 20-30% more African Americans correctily identified as African American, and 60% fewer whites identified as people of color as compared with the status quo BISG method.

Performance on the national PPP loan forgiveness dataset was as follows (comparing ZRP softmax with the BISG method):

African American

Statistic	BISG	ZRP	Pct. Diff
True Positive Rate	0.571	0.700	+23% (F)
True Negative Rate	0.954	0.961	+01% (F)
False Positive Rate	0.046	0.039	-15% (F)
False Negative Rate	0.429	0.300	-30% (F)

Asian American and Pacific Islander

Statistic	BISG	ZRP	Pct. Diff
True Positive Rate	0.683	0.777	+14% (F)
True Negative Rate	0.982	0.977	-01% (U)
False Positive Rate	0.018	0.023	-28% (F)
False Negative Rate	0.317	0.223	-30% (F)

Non-White Hispanic

Statistic	BISG	ZRP	Pct. Diff
True Positive Rate	0.599	0.711	+19% (F)
True Negative Rate	0.979	0.973	-01% (U)
False Positive Rate	0.021	0.027	-29% (F)
False Negative Rate	0.401	0.289	-28% (F)

White, Non-Hispanic

Statistic	BISG	ZRP	Pct. Diff
True Positive Rate	0.758	0.906	+19% (F)
True Negative Rate	0.758	0.741	-02% (U)
False Positive Rate	0.242	0.259	+07% (U)
False Negative Rate	0.241	0.094	-61% (F)

Authors

• Kasey Matthews (Zest AI Lead)

• Austin Li (Harvard T4SG)

• Christien Williams (Schmidt Futures)

• Sean Kamkar (Zest AI)

• Jay Budzik (Zest AI)

Contributing

Contributions are encouraged! For small bug fixes and minor improvements, feel free to just open a PR. For larger changes, please open an issue first so that other contributors can discuss your plan, avoid duplicated work, and ensure it aligns with the goals of the project. Be sure to also follow the Code of Conduct. Thanks!

Maintainers

Maintainers should additionally consult our documentation on releasing. Follow the steps there to push new releases to Pypi and Github releases. With respect to Github releases, we provide new releases to ensure relevant pipelines and look up tables requisite for package download and use are consistently up to date.

Wishlist

Support for the following capabilities is planned:

• …nothing right now! (Got an idea? Submit an issue/PR!)

License

The package is released under the Apache-2.0 License.

Results and Feedback

Generate interesting results with the tool and want to share it or other interesting feedback? Get in touch via abetterway@zest.ai.