ST-DiffTraj: A Spatiotemporal-Aware Diffusion Model for Trajectory Generation

Synthetic traj ectory generation has significant applications in urban planning, mobility analysis, smart city development, etc. Diffusion models excel at generating high-quality synthetic data and are widely used in traj ectory data generation. However, existing diffusion-based trajectory generation models face the following limitations: (1) ignoring temporal sequence information, and (2) inefficient use of conditional guidance. To address these issues, we propose a SpatioTemporal-Aware Diffusion Model for Trajectory Generation, named ST-DiffTraj. Specifically, we design a SpatioTemporal Encoder (STEncoder) to capture correlations between independent and joint spatiotemporal feature distributions by introducing temporal information. Additionally, we develop a noise estimation network by stacking multiple CDAttn-DConv (Condition-Dependent Attention Dilation Convolutional) layers. Each CDAttn-DConv layer integrates attention mechanisms and gated activation units to enhance noise estimation with conditional guidance. Comprehensive experiments on two real-world trajectory datasets demonstrate the superior performance of our model in generating high-quality trajectory data compared to baselines.