Generating geospatial trajectories with incomplete data using graph inverse reinforcement learning

Geographic simulations of human systems often rely on trajectory generation to model agent behavior. This paper addresses trajectory generation under conditions of limited data. We demonstrate how deep learning can reproduce representative urban mobility patterns using two datasets for Jerusalem: a GPS-based travel survey and a GIS dataset of all buildings. To mitigate data mismatch between both datasets, we apply information propagation via multi-dimensional clustering. To exploit the underlying geospatial network structure characterized by transition matrices, we use graph neural networks (GNNs) as reward function approximators and compare them with linear method and multi-layer perceptron as baselines. For the baselines, transition probabilities dominate behavior, while GNNs learn from global data patterns or local neighborhood structures. Results show that GNN-derived rewards correlate more strongly with building floorspace, influencing transition probabilities. We test several initial and terminal state configurations to generate trajectories based on learned policies. In the absence of a universally accepted evaluation method, we assess trajectories using aggregated performance indicators rather than individual trajectory similarity. Findings indicate that GNN-based rewards can improve the realism of simulated trajectories, though further refinement is needed.