Tweet Geolocation: Leveraging location, user and peer signals

• Main Authors: CHONG, Wen-Haw, LIM, Ee Peng
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2017
• Subjects: Query expansion; Tweet geolocation; Collaborative filtering; Databases and Information Systems; Social Media
• Online Access: https://ink.library.smu.edu.sg/sis_research/3956; https://ink.library.smu.edu.sg/context/sis_research/article/4958/viewcontent/TweetGeolocation_2017_afv.pdf
• Institution: Singapore Management University

Which venue is a tweet posted from? We referred this as fine-grained geolocation. To solve this problem effectively, we develop novel techniques to exploit each posting user's content history. This is motivated by our finding that most users do not share their visitation history, but have ample content history from tweet posts. We formulate fine-grained geolocation as a ranking problem whereby given a test tweet, we rank candidate venues. We propose several models that leverage on three types of signals from locations, users and peers. Firstly, the location signals are words that are indicative of venues. We propose a location-indicative weighting scheme to capture this. Next we exploit user signals from each user's content history to enrich the very limited content of their tweets which have been targeted for geolocation. The intuition is that the user's other tweets may have been from the test venue or related venues, thus providing informative words. In this regard, we propose query expansion as the enrichment approach. Finally, we exploit the signals from peer users who have similar content history and thus potentially similar visitation behavior as the users of the test tweets. This suggests collaborative filtering where visitation information is propagated via content similarities. We proposed several models incorporating different combinations of the three signals. Our experiments show that the best model incorporates all three signals. It performs 6% to 40% better than the baselines depending on the metric and dataset.