Mining location and trajectory similarities from human mobility data using natural language processing methods

Keywords: movement trajectory, travel survey, Word2Vec, SemAxis, Word Mover’s Distance
Abstract Type: Paper Abstract

Abstract

Understanding how people move around and participate in various activities in the urban environment is vital to make informed decisions in transportation planning, resource management, and other practices. Advances in location-aware technologies and devices make it possible to conduct travel surveys using mobile devices and automatically record participants’ movement trajectories at the same time. This paper applies methods for natural language processing to analyze such movement trajectories to understand mobility patterns in urban space. Specifically, the paper aims to provide novel insights into the similarity between locations embodied in daily travel and the consequent similarity between individuals’ movements among these locations. To measure similarities between locations, the paper first applies the Word2Vec method to convert locations along each movement trajectory to geo embeddings in vector space and then calculates cosine similarity between each pair of geo embeddings. We find that the similarity between the two locations increases as the trips between them increase. We also associate the cosine distance in vector space with the Euclidean spatial distance in the real world and find that they are correlated but their relationships vary across space. Moreover, we use the SemAxis method to explore whether the geo-embedding of a location can infer land use and population characteristics. Based on results at the location level, we calculate the Word Mover’s Distance to measure the similarity between movement trajectories among locations. We find that trajectory similarities are strongly associated with individuals’ residential locations, suggesting that where people live significantly affects their mobility patterns.