Simultaneous Multiple POI Population Pattern Analysis System with HDP Mixture Regression

Abstract

In recent years, the use of smartphone Global Positioning System (GPS) logs has accelerated the analysis of urban dynamics. Predicting the population of a city is important for understanding the land use patterns of specific areas of interest. The current state-of-the-art predictive model is a variant of bilinear Poisson regression models. It is independently optimized for each point of interest (POI) using the GPS logs captured at that single POI. Thus, it is prone to instability during fine-scale POI analysis. Inspired by the success of topic modeling, in this study, we propose a novel approach based on the hierarchical Dirichlet process mixture regression to capture the relationship between POIs and upgrade the prediction performance. Specifically, the proposed model enables mixture regression for each POI, while the parameters of each regression are shared across the POIs owing to the hierarchical Bayesian property. The empirical study using 32 M GPS logs from mobile phones in Tokyo shows that our model for large-scale finer-mesh analysis outperforms the state-of-the-art models. We also show that our proposed model realizes important applications, such as visualizing the relationship between cities or abnormal population increase during an event.