Identifying the Structure of Cities by Clustering Using a New Similarity Measure Based on Smart Card Data

Identifying the structure of cities has long been studied in urban planning and traffic modeling. This study presents a reliable method that reveals the structure of cities mainly based on clustering analysis using a new similarity measure. Several previous studies have used well-known clustering algorithms in machine learning fields, such as k-means clustering based on temporal mobility patterns of regions, whereas other studies have applied community detection algorithms on networks that depict traffic flows. However, the former does not reflect spatial interactions among places or areas, and the latter groups' places or areas with different land uses into the same cluster. To address these issues in existing approaches, this study proposes a new similarity method that considers not only temporal mobility patterns of areas but also spatial interactions with other areas. Moreover, the study combines spectral clustering repeated several times with hierarchical clustering to obtain a reliable structure that keeps the contiguity of clusters and determine the hierarchy of different areal units. The application of the proposed method to the data for Seoul, South Korea, reveals that the proposed clustering process divides a city into relatively homogeneous areas in terms of land uses. The flow maps based on the clustering results also revealed the spatial interactions between different areas and identify the polycentric structure of Seoul.