Zoning Improvement Plan (ZIP) Codes provide a sub-division of space. Interestingly, the ZIP code area polygons for different data sources do not match, resulting in uncertainty for a range of services that rely on such data. This paper presents a system that employs traditional classification methods to map a given spatial coordinate to a distribution of ZIP-codes using various public available ZIP-code maps as predictors, and using the (not publicly available) United States Postal Service (USPS) map as an authoritative ground truth. We show that large sets of microblog data, from which we extract potential ZIP-codes, can significantly improve classification accuracy despite the noise of such data. The demonstrator allows users to select locations on a map of Orlando, FL, view the resulting distribution of ZIP-codes predicted for this location, compare the results to the ground-truth, and view the microblogs that have enriched the result. A focus will be on showing that the signal present in large, noisy, and 99.99% unrelated microblog data can indeed be used to improve reverse ZIP code geo-coding.

Zoning Improvement Plan (ZIP) Code polygon maps, obtained from different data sources, do not match, thus creating uncertainty in spatial analysis. In this dissertation, we want to combine multiple source of ZIP-code polygon data into a coherent system that maps a given location to a ZIP-code. For this purpose, we want to harness the wisdom of the crowd by combining various polygon map data sets with various sources of volunteered geographical information. The system that we want to build will employ traditional classification methods to map a given spatial coordinate to a distribution of ZIP-codes using the (not publicly available) United States Postal Service (USPS) map as an authoritative ground truth. In our first studies, we train a Naïve Bayes classifier using multiple (publicly available) ZIP Code polygon maps. In addition, we enrich our classification by using a lazy K-Nearest Neighbor classifier to predict the ZIP Codes for a given location using Twitter. Feeding the result of this classifier to the Naïve Bayes classification our experimental evaluation shows an improvement in classification accuracy, compared to using only map data.