Multi-objective representation learning for road networks and trajectories with spatial-temporal fusion and contrastive signals

Modeling and learning representations for road networks and vehicle trajectories are crucial in enabling intelligent transportation systems, with applications ranging from traffic forecasting to many other downstream inference tasks. However, learning effective representations that generalize well across tasks remains challenging due to the heterogeneous nature of spatio-temporal data and limited supervision. In this paper, we propose a unified multi-objective pretraining framework called MRRT, Multi-objective Representation Learning for Road Network and Trajectories, that combines masked trajectory modeling (MTM) with multiple contrastive learning objectives across trajectories, road segments, and spatial contexts. Our model integrates graph attention networks (GAT), spatial CNNs, and transformers with temporal and positional encoding, allowing us to capture structural and contextual dependencies in urban mobility. By leveraging grid-structured and graph-structured data, along with spatiotemporal dynamics, our model effectively captures diverse road and trajectory characteristics. To enhance robustness, we design trajectory-specific data augmentations and contrastive heads for trajectory-to-trajectory, trajectory-to-node, and node-to-node alignment. Additionally, we design an adaptive negative sampling strategy to further enhance the contrastive learning. We evaluate our approach in various downstream tasks based on trajectory and road, including travel time estimation, speed inference, and similarity search. Extensive experiments demonstrate that our method consistently outperforms prior baselines and ablated variants, validating the effectiveness of our multiobjective design.