Identifying user geolocation with Hierarchical Graph Neural Networks and explainable fusion

Determining user geolocation from social media data is essential in various location-based applications — improved transportation/supply management, through providing personalized services and targeted marketing, to better overall experiences. Previous methods rely on the similarity of posting content neighboring nodes for geolocation, which suffer problems of: (1) position-agnostic network representation learning, impedes performance their prediction accuracy; (2) noisy unstable relation fusion due flat graph embedding employed. This work presents Hierarchical Graph Neural Networks (HGNN) – a novel methodology location-aware collaborative user-aspect location prediction. It incorporates geographical information users clustering effect regions can capture topological relations while preserving relative positions. By encoding structure features with hierarchical HGNN primarily alleviate problem signal fusion. We further design mechanism bridge connections between individual clusters, not only leverages isolated that are useless previous but also captures unlabeled labeled subgraphs. Furthermore, we introduce robust statistics method interpret behavior our model by identifying importance samples when predicting locations users. provides meaningful explanations behaviors outputs, overcoming drawbacks approaches treat as “black-box” modeling lacking interpretability. Comprehensive evaluations real-world Twitter datasets verify proposed model’s superior its ability results.