Geolocation using GAT with Multiview Learning

Information in social networks plays an important role in many fields such as event detection, disaster warning, etc. However, due to the lack of geographic metadata, the information is often unusable. Therefore, the geolocation using social network data has gradually become a hot research topic. Existing methods mainly use textual contents, and thus poorly exploit the available data, especially the hidden information in the link. To address this issue, we propose two Multiview learning models, M-GAT and M-GCN, based on the Graph Attention and Graph Convolution Network to fuse both the text and link information. By extracting the text features from multiple angles to extend the feature space, our models achieve the best results on the baseline dataset. The visual display of representations collected from a hidden layer illustrates the validity of our models. Experiments on different feature combination show the effectiveness of our proposal.