Heterogeneous Graph Attention Network for User Geolocation

• Published in: PRICAI 2021: Trends in Artificial Intelligence pp. 433 - 447
• Main Authors: Zhang, Xuan, Lin, FuQiang, Dong, DiWen, Chen, WangQun, Liu, Bo
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Attention mechanism; Heterogeneous graph; User geolocation
• ISBN: 9783030893620; 3030893626
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-89363-7_33

Identifying the geographic location of online social media users, also known as User Geolocation (UG), plays an essential part in many Internet application services. One main challenge is the scarcity of users’ public geographic information. To overcome it, most works focus on user geolocation prediction with posts and interactions on social media. However, they do not consider the distinction of variant social connections, which may impair the performance of the UG models. To address this issue, we propose a multi-view model, Heterogeneous graph Attention networks for user Geolocation (HAG), which introduces the attention mechanism to mine valuable cues in social networks and text contexts jointly. In the network module, we creatively apply a heterogeneous graph to model various social interactions and introduce a heterogeneous graph attention network to learn network structure information. In the text module, we propose a context attention network to extract geo-related text information. Extensive experiments conducted on three Twitter datasets show that HAG achieves state-of-the-art performance compared to strong baselines.