Spatially explicit knowledge in geo-embeddings: Interpreting location representation derived from human movement trajectories

The Word2Vec model, originally developed for natural language processing, has been widely used to encode geographical locations such that embedded spatial information can be incorporated into machine learning and artificial intelligence models. These geo-embeddings of locations derived from human movement trajectories can capture the unique roles of locations in facilitating our everyday lives. However, most studies utilize the derived geo-embeddings for downstream tasks, while only a few studies have interpreted geo-embeddings using intrinsic approaches and explored their implications explicitly. Our study addresses this gap by evaluating whether the geo-embeddings generated by human movement trajectories can characterize the interconnectivity among locations, and how such knowledge can be further discerned from spatially explicit perspectives such as distance and spatial interaction flows. We develop an evaluation framework that considers two location representations and multiple (dis)similarity metrics regarding mobility flow patterns to examine the effectiveness of geo-embeddings in capturing spatially explicit knowledge. Large-scale mobile positioning data collected in the Twin Cities metropolitan area, Minnesota, U.S. is used as a case study to generate geo-embeddings for census tracts. Results suggest that the Word2Vec geo-embeddings can capture the distance decay effect of location interactions and spatial structure of sociodemographics. The spatial distribution of sociodemographic semantics derived from geo-embedding displays a similar pattern to the census income and race data. Moreover, the geo-embeddings can facilitate meaningful learning outcomes for enhanced community and sociodemographic studies.