<p>Recently, data-driven traffic assignment solutions have been developed using different graph-based neural network architectures. Although analytical and simulation-based solutions rely on the shortest paths to find optimal routes between nodes in the network, shortest path information is not included in data-driven solutions. In this paper, we develop a novel graph transformer model that uses shortest path information as edge features to generate data-driven traffic assignment solutions. The model utilizes a dynamic attention module to effectively capture the impact of links that are present in shortest paths between node pairs and enhances the model’s ability to learn the varying importance of different links in the network. We run numerical experiments on two networks (Sioux Falls and Eastern-Massachusetts) and show that adding shortest path information outperforms state-of-the-art neural network models in predicting link flows (9.65% and 3.92% improvement of RMSE for Sioux Falls and Eastern-Massachusetts networks, respectively compared to a graph transformer model without shortest path information). We also develop a transfer learning method and test it using a model trained on Sioux Falls network to predict link flows of Eastern-Massachusetts network and find that shortest path information enhances the generalization capability of the model to different networks. Finally, we present an analysis of whether the predicted flows are close to equilibrium flows. Thus, by improving prediction accuracy and generalization capability, the developed approach based on graph transformer and shortest path information enhances the current solutions of data-driven traffic assignment problems.</p>