BERT-based Dynamic Clustering of Subway Stations Using Flow Information

Man Li

Abstract: The increasing volume and complexity of metro systems urge the transportation agencies to gain more knowledge of the mobility patterns of metro stations to improve public service, adjust future planning and even reconstruct the network. Therefore understanding the dynamic functions of metro stations becomes essential. In this work, we propose a BERT-based feature extraction framework to capture dynamic mobility patterns for metro stations time to time. Specifically, we adopt the flow counts gained from Automated Fare Collection (AFC) systems in every time interval as the representation of current mobility pattern and design the flow matrices for each station. The proposed feature extraction framework is implemented to learn the latent features and once the latent semantics are obtained, we apply affinity propagation clustering algorithm to segment subway stations into different clusters. Based on the dynamic clustering result, further work can be done like group flow prediction and anomaly detection.

External IDs:dblp:conf/icde/Li21a