Encoding Urban Trajectory As A Language: Deep Learning Insights For Human Mobility Pattern

Rapid advancements in deep learning technology have shown great promise in helping us better understand the spatio-temporal characteristics of human mobility in urban areas. There exist two main approaches to spatial deep learning models for urban space - a convolutional neural network (CNN) which originated from visual data like satellite image, and a graph convolutional network (GCN) which is based on the urban topologies such as road network and regional boundaries. However, compared to language-based models that have recently achieved notable success, deep learning models for urban space still need further development. In this study, we propose a novel approach that addresses the trajectories of a trip as sentences of a language and adapts techniques like word embedding from natural language processing to gain insights into human mobility patterns in urban areas. Our approach involves processing sequences of spatial units that are generated by a human agent's trajectory, treating them as akin to word sequences in a language. Specifically, we represent individual trajectories as sequences of spatial vector units using 50×50 meters grid cells to divide the urban area. This representation captures the spatio-temporal changes of the trip, and enables us to employ natural language processing techniques, such as word embeddings and attention mechanisms, to analyze the urban trajectory sequences. Additionally, we leverage word embedding models from language processing to acquire compressed representations of the trajectory. These compressed representations contain richer information about the features, while minimizing the computational load.