Identifying Spatiotemporal Patterns of Transfers between Subways and Buses through Two-Step Clustering

Identifying spatiotemporal patterns of transfers between subways and buses is crucial for promoting the integrated development of urban public transportation systems. However, existing studies focused on the spatiotemporal correlation of single-mode travel, ignoring heterogeneity in the spatiotemporal distribution of transfers during the combined travel process of subways and buses. Leveraging large-scale smart card data enabled by advances of information and communication, this study proposes a novel two-step clustering approach to identify spatiotemporal patterns of transfers between subways and buses. Firstly, based on the spatial distribution characteristics of origin and destination points at transfer stations, the comprehensive transfer similarity (CTS) metric is proposed to measure the spatial similarity of combined travels across different time intervals and transfer stations. Secondly, a CTS-based clustering framework utilizing the fast search and find of density peaks (CTS-CFSFDP) algorithm is proposed. This framework tackles the problem of defining the local spatial density and the minimum spatial distance to higher-density stations when identifying spatial patterns at each transfer station across different time intervals. Finally, the sequence of spatial pattern changes at each station over a single day is used as the basic unit for the affinity propagation (AP) algorithm to conduct further clustering and identify the time-series patterns of spatial distributions at these stations. Using 1 month’s public transport travel chain data from Beijing, we identify seven distinct time-series patterns derived from five spatial patterns of transfer travel. Within these different spatial patterns of transfer travel, the bus network fulfills distinct roles: acting as a connection, distribution, or accessibility in relation to the subway network. Additionally, through a comparative analysis of enrichment factors (EF) associated with diverse points of interest (POIs) across the identified time-series patterns, the consistency between the clustering results and land-use function analysis of POIs is verified.