Large-Scale, Fine-Grained, Spatial, and Temporal Analysis, and Prediction of Mobile Phone Users’ Distributions Based upon a Convolution Long Short-Term Model

• Published in: Sensors (Basel, Switzerland) Vol. 19; no. 9; p. 2156
• Main Authors: Zhang, Guangyuan, Rui, Xiaoping, Poslad, Stefan, Song, Xianfeng, Fan, Yonglei, Ma, Zixiang
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 09.05.2019; MDPI
• Subjects: autoregressive moving average (ARMA); Big Data; Cellular telephones; Censuses; convolution LSTM (ConvLSTM); Data mining; Deep learning; Global positioning systems; GPS; International conferences; kernel density estimation (KDE); long short-term memory (LSTM); mobile phone data; Neural networks; Population density; Sensors; Short term; spatial–temporal data analysis and prediction; Teaching methods; Time series; Traffic congestion; Traffic flow; Wireless networks; United Kingdom > UK; United States > US; China; deep learning; population density; convolution LSTM (ConvLSTM); kernel density estimation (KDE); long short-term memory (LSTM); spatial–temporal data analysis and prediction; mobile phone data; autoregressive moving average (ARMA)
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s19092156

Accurate and timely estimations of large-scale population distributions are a valuable input for social geography and economic research and for policy-making. The most popular large-scale method to calculate such estimations uses mobile phone data. We propose a novel method, firstly based upon using a kernel density estimation (KDE) to estimate dynamic mobile phone users’ distributions at a two-hourly scale temporal resolution. Secondly, a convolutional long short-term memory (ConvLSTM) model was used in our study to predict mobile phone users’ spatial and temporal distributions for the first time at such a fine-grained temporal resolution. The evaluation results show that the predicted people’s mobility derived from the mobile phone users’ density correlates much better with the actual density, both temporally and spatially, as compared to traditional methods such as time-series prediction, autoregressive moving average model (ARMA), and LSTM.