Language independent big-data system for the prediction of user location on Twitter

Abstract

Social media interactions have become increasingly important in today's world. A survey conducted in 2014 among adult Americans found that a majority of those surveyed use at least one social media site. Twitter, in particular, serves 310 million active users on a monthly basis, and thousands of tweets are published every second. The public nature of this data makes it a prime candidate for data mining. Twitter users publish 140-character long messages and have the ability to geo-tag these tweets using a variety of methods: GPS coordinates, IP geolocation and user-declared location. However, few users disclose their location, only between 1% and 3% of users provide location data, according to our empirical findings. In this article, we aim to aggregate information from different sources to provide an estimation on the location of any Twitter user. We use an hybrid approach, using techniques in the fields of Natural Language Processing and network theory. Tests have been conducted on two datasets, inferring the location of each individual user and then comparing it against the actual known location of users with geolocation information. The estimation error is the distance in kilometers between the estimation and the actual location. Furthermore, there is a comparison of the relative average error per country, to account for difference in country sizes. Our results improve those presented in different researches in the literature. Our research has as feature to be independent of the language used by the user, while most of works in the literature use just one language or a reduced set of languages. The article also showcases the evolution of our estimation approach and the impact that the modifications had on the results.