Estimating locations of social media content through a graph-based link prediction

The increasing availability of GPS-enabled devices and social media platforms has led to an increasing interest in mining geolocated content. However, our understanding of the role played by social media in the social construction of place has been limited by the fact that only a small percentage of social media posts are geolocated. Spatio-temporal modelling research in this field so far has mainly focused on analysing the behaviour of single users and predicting user movement patterns based on posts and check-in activities. In this paper, we focus instead on harnessing the dynamics of overall content production from multiple users in a single place to estimate the location of new non-geotagged content. Our proposed location prediction framework uses a variational graph autoencoder, and it allows us to estimate the geolocations of posts based on the semantic understandings of their contents and their topological structure.