カンファレンス (国際) Simultaneous multiple POI population pattern analysis system with HDP mixture regression
Yuta Hayakawa (Tokyo Institute of Technology), Kota Tsubouchi, Masamichi Shimosaka (Tokyo Institute of Technology)
the 25th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD2021)
The spread of mobile phones in recent years has accelerated the analysis of urban dynamics by allowing the use of the Global Positioning System (GPS) logs of mobile phones. In particular, predicting the population in a city has great im- portance for understanding land use patterns of certain areas of interest. The state-of-the-art model, a variant of bilinear Poisson regression models, related to each point of interest (POI) is independently optimized using the GPS logs cap- tured in a single POI; thus it is prone to be unstable for finer scale POI analysis. Inspired by the success of topic modeling and collaborative filtering, we consider how to capture the relationship between POIs to upgrade the prediction per- formance. In this paper, we propose a new approach based on hierarchical Dirichlet process (HDP) mixture regression sharing the dataset across POIs to improve the accuracy of the prediction and also the quality of the urban dynamics analysis system based on our proposed model. Specifically, the proposed model results in mixture regression per POI, whereas the parameters of each regression are shared across the POIs thanks to the hierarchical Bayesian property. We also show our proposed model for realizing the prominent applications needed in industry, for such as visualizing the relationship between cities or an abnormal increase of popu- lation due to an event. The empirical study using 32 M GPS logs from mobile phones in Tokyo, Japan shows that our model for large-scale finer-mesh analysis outperforms the state-of-the-art models.