文章基本信息

标题：Fine-Grained Urban Functional Region Identification via Mobile App Usage Data
本地全文：下载
作者：Lei Deng ; Hangyu Hu
期刊名称：Mobile Information Systems
印刷版ISSN：1574-017X
出版年度：2022
卷号：2022
DOI：10.1155/2022/6434598
语种：English
出版社：Hindawi Publishing Corporation
摘要：Understanding fine-grained urban function for different regions is essential for both city managers and residents in terms of strategy design, tourism recommendation, business site selection, etc. A huge amount of data from the mobile network in the past several years provides the possibility for fine-grained urban function identification since it provides the opportunity to extract useful information about urban functions. However, challenges remain: (i) there is no prior knowledge about the existence of App usage patterns relating to urban functional regions; (ii) collected data are very noisy and data from different cellular towers have different noise levels. Therefore, it is difficult to extract unique patterns to identify urban functional regions. This article proposes a fine-grained urban functional region identification system, which utilizes mobile App usage data from cellular towers. To address challenge (i), we first extract three key variables for each cellular tower, App number, user number, and traffic. Then, we design a Davies–Bouldin index (DBI)-based filtering method to automatically select the most distinguishable features for multiclassification. To address challenge (ii), we first reduce cellular tower level noise by designing a clustering-based method to select the most representative cellular tower data. The data from these cellular towers share similar patterns for the same urban functional region and different patterns between different urban functional regions. Then, we reduce feature level noise by designing a Fourier transform-based method to reconstruct the features with several key frequency components, which preserves the most important information and removes the unnecessary noise. We evaluate our system and selected features with three representative supervised learning models, all of which achieve more than 95% classification accuracy.