Social user geolocation based on K-medoids and Gaussian Kernel graph attention network

Accurate user location information is crucial for many location-based network services. However, existing social user geolocation methods using fixed-grid partitioning fail to accurately locate users in rural areas. Additionally, these methods ignore the distances between node features, leading to deviations in users' location features and thus reducing the accuracy of user localization. To address these challenges, this paper proposes a novel social user geolocation method (KMKGAT) based on k-medoids and Gaussian kernel graph attention network. Specifically, KMKGAT employs an anti-noise k-medoids algorithm to cluster user locations, ensuring precise clustering of geographically adjacent users. At the same time, by introducing parameterized Gaussian kernel functions into the graph attention network, KMKGAT learns location-enhanced user features from text-featured social networks, thereby alleviating the problem of location feature deviation. Extensive experiments are conducted on three public Twitter datasets. The experimental results show that the proposed method is superior to the state-of-the-art baselines.