Domain-Knowledge Enhanced GANs for High-Quality Trajectory Generation

Realistically simulating human mobility and generating high-quality trajectories are essential for location-based applications like epidemic analysis, traffic management, and location privacy. Current trajectory generation techniques that are free of models usually learn knowledge directly from real mobility data. However, they encounter challenges in generating high-quality trajectories due to the unpredictable transition patterns and intricate periodicity regularities inherent in human movement. Furthermore, model-free techniques rely heavily on autoregressive paradigms, which are susceptible to issue such as error accumulation. In order to address these challenges, we propose Domain-Knowledge Enhanced Generative Adversarial Network (DKE-GAN), a model-free approach that integrates domain knowledge of human mobility with model-free learning paradigm to generate high-quality human mobility data. Additionally, we tackle the error accumulation issue by integrating reinforcement learning into the discrimination stage. The discriminator here gradually supplies augmented feedback, incorporating sequence generation, mobility regularity awareness, and mobility yaw rewards, to offer comprehensive guidance for enhancing the generator's performance. Extensive experiments conducted on two real-world mobility datasets show that our framework outperforms five state-of-the-art baselines, significantly improving the simulation of human mobility data.