Predicting the location of users on Twitter from low density graphs

Abstract

In this paper we present an algorithm for the estimation of geographic locations of Twitter users. The algorithm is based on the graph representation of communication patterns on Twitter and employs an effective clustering algorithm for estimating locations of users from their connections in the communication graph. While using a graph based approach to estimate geo-location for Twitter users is not new, most of the existing methods require very dense networks which generally require collecting data over a considerable period of time. In this paper, we present a new method that is able to achieve good accuracy and coverage for geolocation estimation of Twitter users using a much less dense graph and only requiring a little over one month's worth of data. The approach is based on label propagation and a new inference location method that uses a clustering of geographical points. We test it on two versions of the Twitter graph built from data in June-July of 2014 and December 2014-January 2015. Analysis of the results demonstrates high accuracy — 65% of geo-labeled users have distance error between true and inferred locations less than 50 km. We were able to geolocate up 87% of users from our datasets. When comparing results between the two graphs, we show that around 40% of users move more than 25 km over the 6 month period.