Predicting trajectory destinations based on diffusion model integrating spatiotemporal features and urban contexts

Accurately predicting travel destinations in a city from partial trajectories is crucial for many location-based services. This task faces two challenges: data sparsity and the complex probability distribution of destinations. Previous studies neglected the fusion of spatiotemporal features and urban contexts, which limited their effectiveness in characterizing trajectory patterns and exacerbated data sparsity issues. Most methods either predicted a single destination or the probability of each predefined sub-region, but they struggled with the modifiable areal unit problem. To address these problems, we propose an approach based on the diffusion probabilistic model, comprising a trajectory encoder and a diffusion decoder. The trajectory encoder extracts four embeddings to represent spatiotemporal features and urban contexts, integrating them using an attention mechanism to mitigate data sparsity. The diffusion decoder recovers the complex destination probability distribution and predicts multiple candidate destinations without predefined regional units, effectively addressing the modifiable areal unit problem. Experiments on driving trajectories in Shenzhen, China, show an average improvement of 10.80% across four metrics compared to state-of-the-art methods. We also observed the influence of trajectory length and origin location on prediction accuracy. The results validate the importance of modeling destination probability distributions, as well as integrating spatiotemporal features and urban contexts.