Location Prediction in Social Media Based on Contents and Graphs

The location information in social media often provides insight on the repercussions of events and is becoming more and more vital to the applications such as localized search, local news recommendation and emergency detection. However, for the security and privacy concerns, most users in social media are unwilling to publish their location, and they may post non-existing places or submit unspecific place names. We focus on analysis post content and social graph to predict users' resident location in city-level level. Based on a fraction of the users with known locations in their profiles, we propose a new filter to identify location sensitive words in posts content. We also expand the social relationships of users and combine these two approaches to propose a mixture probabilistic model to estimate user location. The experimental results on a large scale of dataset crawled from Tencent weibo demonstrate that our approach achieve an accuracy of 60.2% in city level and outperform state-of-the-art approaches.