A0913
Title: Detecting changes in the spatiotemporal pattern of bike-sharing: A change-point topic model
Authors: Xian Chen - Shanghai University (China) [presenting]
Kun Huang - Texas AM University (China)
Hai Jiang - Tsinghua University (China)
Abstract: Detecting latent changes in the spatiotemporal pattern of bike sharing is critical to recognizing the impacts of various exogenous factors and significant events. Although extensive studies have investigated the spatiotemporal pattern dynamics in transportation systems, research that explicitly models the changing regularity and captures its latent changes is very scarce. To fill this research gap, a change-point topic model is developed that incorporates multinomial logit models into multi-dimensional latent Dirichlet allocation to decompose the spatiotemporal pattern into a mixture of activity patterns; meanwhile, the model captures the changing regularity of activity prevalence and its latent changes by integrating Dirichlet multinomial regression and Dirichlet process hidden Markov models. The parameters of the model based on collapsed Gibbs sampling are estimated. Numerical experiments are conducted using publicly available bike-sharing trip records collected in New York. Results show that the model successfully distinguishes several meaningful activity categories, such as commuting. Furthermore, the detected pattern changes offer several insights into the impacts of associated events. The model also improves the goodness of fit and predictive performance for the spatiotemporal attributes of trips.
