Identifying subway commuters travel patterns using traffic smart card data: A topic model

Abstract

The paper presents a novel approach using Hierarchical Dirichlet Processes (HDP) integrated with K-means clustering to analyze public transit commuting behaviors using smartcard and POI data. The HDP, an unsupervised model, is designed to discern travel activities, however, little is done for this purpose. Our study proposed representing each trip using four features (duration, date, arrival time, and station type classified using POI-data) as inputs to the HDP model, which outputs the identification of specific activities such as home, work, and leisure. A comparison to other methods including trip frequency, activity duration, and Hidden Markov models demonstrates that our approach offers superior fit, as evidenced by lower perplexity and higher similarity metrics. To further refine the classification of commuting behaviors, we applied a two-step clustering algorithm that considers features such as regularity, temporality, and spatiality, resulting in the identification of strong and weak commuting behavior patterns. This classification provides urban planners with insights into the spatiotemporal characteristics of travelers in urban rail transit systems, thereby supporting more effective urban planning.