An Unsupervised Approach to Identify Location Based on the Content of User's Tweet History

摘要

We propose and evaluate an unsupervised approach to identify the location of a user purely based on tweet history of that user. We combine the location references from tweets of a user with gazetteers like DBPedia to identify the geolocation of that user at a city level. This can be used for location based personalization services like targeted advertisements, recommendations and services on a finer level. In this paper, we use convex hull and k-center clustering, to identify the location of a user at a city level. The main contributions of this paper are: (i) reliability on just the contents of a tweet, without the need for manual intervention or training data; (ii) a novel approach to handle ambiguous location entries; and (iii) a computational geometric solution to narrow down the location of the user from a set of points corresponding to location references. Experimental results show that the system is able to identify a location for each user with high accuracy within a tolerance range. We also study the effect of tolerance on accuracy and average error distance.