CondTraj-GAN: Conditional Sequential GAN for Generating Synthetic Vehicle Trajectories.

While the ever-increasing amount of available data has enabled complex machine learning algorithms in various application areas, maintaining data privacy has become more and more critical. This is especially true for mobility data. In nearly all cases, mobility data is personal and therefore the drivers’ privacy needs to be protected. However, mobility data is particularly hard to anonymize, hindering its use in machine learning algorithms to its full potential. In this paper, we address these challenges by generating synthetic vehicle trajectories that are not subject to personal data protection but have the same statistical characteristics as the originals. We present CondTraj-GAN– Cond itional Traj ectory G enerative A dversarial N etwork. – a novel end-to-end framework to generate entirely synthetic vehicle trajectories. We introduce a specialized training and inference procedure that enables the application of GANs to discrete trajectory data conditioned on their sequence length. We demonstrate the data utility of the synthetic trajectories by comparing their spatial characteristics with the original dataset. Finally, our evaluation shows that CondTraj-GAN reliably outperforms state-of-the-art trajectory generation baselines.