Retrieval-based human trajectory generation

We considered the advantages and limitations of directly using the generative model for human mobility estimation and considered improving two of these limitations. The first limitation is that when we generate new virtual data directly using the generative model, the newly generated trajectory data does not comply with geographic information constraints. That is, the car could appear in an area outside the road network. The second limitation is that it is difficult for us to quantitatively measure the authenticity of the newly generated virtual trajectory data. We naturally think that we can use map matching to match the newly generated trajectory to the road network regarding the first limitation. However, we have noticed that such postprocessing will change the citywide human mobility pattern. Therefore, we first used the shortest distance with the map matching method to conduct a simple experiment to measure the change in the citywide human mobility pattern of the postprocessing virtual trajectory. The experimental results show that the citywide human mobility pattern represented by the postprocessing trajectory data has probably changed by more than 20%. Then, from the perspective of trajectory similarity, we use the retrieval idea to construct a retrieval-based human mobility estimation model. In this way, we can avoid the human mobility pattern change brought by map matching as postprocessing and avoid the problem of quantifying the authenticity of the newly generated virtual trajectory. We first use the deep learning model to convert complex trajectories into hidden space distributions. We then use the distance between the hidden space distributions corresponding to different trajectories to complete a quick search with the k-d tree technique. In the experiment, we compared the deep learning model with the traditional trajectory similarity method. We found that the deep learning model, especially based on the two-way Long Short-Term Memory and VAE model, obtained the best results. The limitation of the retrieval-based model is that we need a vast historical trajectory database, and we do not know how to estimate the appropriate weight of each observed trajectory in citywide human mobility.