Landmark-based user location inference in social media

Abstract

Location profiles of user accounts in social media can be utilized for various applications, such as disaster warnings and location-aware recommendations. In this paper, we propose a scheme to infer users' home locations in social media. A large portion of existing studies assume that connected users (i.e., friends) in social graphs are located in close proximity. Although this assumption holds for some fraction of connected pairs, sometimes connected pairs live far from each other. To address this issue, we introduce a novel concept of landmarks, which are defined as users with a lot of friends who live in a small region. Landmarks have desirable features to infer users' home locations such as providing strong clues and allowing the locations of numerous users to be inferred using a small number of landmarks. Based on this concept, we propose a landmark mixture model (LMM) to infer users' location. The experimental results using a large-scale Twitter dataset show that our method improves the accuracy of the state-of-the-art method by about 27%.