Road Network-based Vehicle Trajectory Imputation via Generative Adversarial Network

Dense trajectories are important to many location-based service (LBS) applications, e.g., traffic prediction.Existing works often fail to ensure that imputed vehicle trajectories perfectly align with road networks.The two-stage approach involving first trajectory imputation and next map-matching suffers from high computation cost and may not leverage the inherent relationship between trajectories data and road networks.To address the issue, in this paper, we develop a generative adversarial networks (GAN)-based imputation framework by leveraging road networks to fill missing points, and train the GAN model by policy-based reinforcement learning (RL) for better integrity and accuracy.Our preliminary evaluation on two real datasets demonstrates that our work outperforms state-ofthe-art with better effectiveness and efficiency.