A graph convolutional network for social user geolocation based on topics and social relationships

Geolocation inference is a critical issue for location-based applications, including natural disaster monitoring, flu trend prediction, and targeted advertising. However, social networks often provide sparse and unreliable location information, posing significant challenges for existing location inference methods. To address these limitations, we introduce a novel home location inference model, the Social Relationship-based Graph Convolutional Network (SRGCN), which integrates textual topics and graph neural networks (GNNs) to effectively capture social relationships. SRGCN first extracts hybrid features from text using TF-IDF to generate textual feature vectors and constructs an initial social relationship graph based on user mentions. Then, a topic model is employed to establish connections among isolated users, thereby enhancing the graph structure. Finally, we leverage graph convolutional neural networks to process the SRG and jointly model textual features and network structure for accurate geolocation prediction. Using the GeoText dataset, we assess the impact of topic similarity thresholds on performance and graph size. Results show that SRGCN aggregates same-class user nodes, increases the number of locatable users, and preserves association diversity. It outperforms existing methods like MENET and GCN4Geo in error metrics and accuracy, achieving a 1.0 % accuracy improvement and a 2.9 % average error reduction over GCN4Geo. By linking isolated users through topic similarity, SRGCN enriches the social relationship graph and significantly improves location inference accuracy. • SRGCN integrates textual topics and GNNs for accurate geolocation prediction. • Tackles challenges in user feature analysis and social relationship representation. • Topic model connects isolated users, enriching the social relationship graph.