Algorithmic Geolocation of Brazilian Polling Places Using GIS Census Data

• In Event: Geography and Spatial Analysis

Abstract

In this paper, we develop an algorithm to determine the latitude and longitude of the polling places employed in Brazil’s elections between 2002 and 2020, using as input their addresses and as auxiliary data, 3 unrelated well-known census datasets that contain geolocated Brazilian schools, rural properties or urban street blocks. Our algorithm works in two phases. First, we propose a novel fuzzy name matching procedure (based on an iterative progressive substring matching) that is capable of correctly matching polling place establishment names to establishment names present in those other census datasets with near 99% accuracy (assessed via MTurk sampled checking). Importing geolocation from those other official datasets through that procedure is sufficient to solve up to 80% of the polling places’ latitude and longitude. In the second phase, the remaining 20% cases are addressed by using a Deep Neural Network to choose the best geographical coordinate candidate, trained on the successful classifications from phase one and from ground truth data described next.

To verify the accuracy of our generated geolocations, we use as a ground truth benchmark the official geolocated polling place addresses that we gathered for around half of the 2020 polling places.
Brazil’s national electoral authority, the Supreme Electoral Court (TSE), provides physical addresses of all the polling places used since at least 2002. Until 2018, just a very small fraction of those were officially geolocated by TSE. In 2018, they were 28% and, in 2020, 38%. In direct contact with the 27 State Electoral Courts (TREs), we were able to find additional unpublished geolocations officially performed by the TREs for extra 12% of the country’s current polling places, to a total of around half of the 2020 places officially geolocated by some electoral authority. While we apply our proposed algorithm to all polling places from all years, we use those 50% from 2020 as a ground truth benchmark. Validated geolocations from the first phase of our approach had a median error of just 29 meters from the ground truth . The ones from the second phase showed a median error of around 100 meters.

To better understand our approach, consider the following. From the country’s current 94933 polling places, around 2/3 are schools. The rest are a mix between public heath buildings, other public administration building, some churches, indigenous tribes and, rarely, private property. TSE and the TREs do provide the names of the establishments being used as polling places. Fortunately, Brazil has unrelated census data where many types of establishments have some level of geocoding information, which can be accurately matched by name with our algorithm. One of those auxiliary datasets come from the country’s School Census, where close to 150 thousand schools have been geolocated. The other two census datasets comprise all the countryś public and private physical address, as used in the 2010 national census, and all rural properties from the 2017 agricultural census. In them, rural addresses and properties were officially geolocated. While urban addresses were not, they were linked to each of the country’s specific street blocks – and centroid for those are available.

Therefore, a really accurate name matching is capable of using those datasets to find the correct coordinates of dozens of thousands of Brazillian polling places. This novel name matching relies on iteratively evaluating substrings (separated by words), from left to right, contained within each polling place name versus the names contained in the other datasets. This accounts for the fact that the nearly always, the differences in name writing for a same establishment are simply due to different word orders, abbreviations and the alike. Hence, substring matching is capable of reliably solving the issue. In the paper, the algorithm is described in detail, together with its pseudo-code. In order to verify the accuracy of the name string matching in itself, we resorted to Amazon MTurk to enlist Portuguese speaking users to verify randomly selected pairs of matched names. Around 99% of them were classified as unambiguous matching.

Naturally, however, there are residual cases that cannot be reliably solved for other reasons: for instance, when some establishments’ names are simply lacking from the auxiliary datasets. To handle those, in our second phase we look for traditional fuzzy string matchings in establishment names, addresses and neighborhood, from each of the three auxiliary datasets. We also use Google Maps and ESRI ArcMap address geolocation APIs to get corresponding coordinates. Then, we train a Deep Neural Network (using ground truth data and successful pairings from phase 1) on those many less-than-ideal matchings to adjudicate, for each polling place, which should be considered the best source for geolocation. As previously specified, validation using ground truth data proves excellent.

Author

• Fabricio Vasselai, University of Michigan