<p>Graph representation learning has shown promising potential for improving global navigation satellite system (GNSS) positioning accuracy, which can analyze unstructured multi-constellation satellite measurements to extract environmental features. However, dynamic urban environments have complex satellite distribution changes because of occlusion from the surrounding environment, and the current methods of stacking Graph neural networks (GNNs) with a message-passing mechanism struggle to accommodate this complex satellite graph structure, ultimately degrading positioning accuracy. In this paper, a transformer based on graph-structure-learning (STL-GT) correction method is proposed to improve positioning accuracy. Specifically, to accommodate differences in satellite distribution across urban environments, the feature set of satellite measurements across multiple constellations is expanded, and the graph structure is augmented with position coding to differentiate between satellite nodes at different locations. To extract environmental features efficiently, a graph transformer network based on subgraph extraction is developed, integrating both local and global structural information of satellite observation graphs to generate satellite graph representations with environmental differences. Finally, we improve the model with a new robust training strategy by introducing disturbance factors in the model optimization to simulate deviations and disturbances in the data distribution, thereby enhancing the accuracy and stability of the STL-GT. Experimental results demonstrate that STL-GT significantly improves GNSS positioning performance in urban environments, e.g., achieving a 49% accuracy improvement in the urban trajectory of GSDC compared with the WLS + KF baseline, and a 17% improvement in positioning performance over the state-of-the-art graph learning-based methods in the GZGNSS data set.</p>