Mining Activity Chains From Large-scale Mobile Phone Location Data

An individual's activity chain generally refers to a series of daily activities with characteristics such as activity order,time,location,and purpose.Understanding residents' activity chains with both spatial and temporal characteristics provides critical support for human mobility research and applications in urban planning,transportation,public health and many other related fields.Traditional approaches to collecting activity chain information mainly depend on travel surveys,which often take up a large amount of resources but lack of instantaneity and continuity.Recent information and communication technologies(ICT)ubiquitously generate individual trajectory data,thus offering the possibility of continuously perceiving human activity chains with large sample sizes and low cost.One of the commonly used trajectory datasets is mobile phone location data,which has been suggested quite useful for mining activity patterns in urban areas.Detecting individual stay areas and inferring activity purpose are important and basic steps in many studies based on mobile phone location data.However,it still remains challenges on robustly detecting an individual's activity locations and time from the messy and noisy mobile phone location data and effectively inferring activity purposes overcoming the uncertainty in space and time embedded within this data.Particularly,the sparse spatial and temporal resolution of raw mobile phone location data and data noise caused by location oscillation and location drift introduce great challenges in effectively detecting individual stay areas from raw mobile phone location data.What's more,lack of semantic information such as activity purpose,researchers can not achieve more knowledge from raw mobile phone location data.Although some recent machine learning models seem to perform well on activity purpose inference,most of them work as black boxes for domain experts,which fail to offer insights to understand the interactions between built environment and residents' daily activities from both the spatial and temporal perspectives.To address the above challenges,this study proposes a white box method to mine human activity chains from large-scale mobile phone location data by integrating both the spatial and temporal feature of daily activities.Specifically,we firstly design a stay areas detection algorithm to find out when and where a mobile phone user stays for certain activities from raw mobile phone location data.This proposed algorithm combines the properties of a moving window filter with an incremental clustering algorithm together to overcome the weakness of only using constraints in travel distance and stay duration in most existing algorithms of activity stop detection.Then through integrating spatial features from surrounding built environment and temporal features from the changing relationships between adjacent activities via time,we put forward a mixed model based on linear interpolation to infer activity purposes.These spatial and temporal features are separately measured by spatial statistics and transition probabilities based on Markov property,and then are linearly combined together to generate a probability array based on which we can infer the activity purpose by using Monte Carlo simulation.Finally,we construct an individual's activity chains with activity order,location,time,and purpose.Using a large-scale mobile phone location dataset of Shenzhen City,this study demonstrates the effectiveness of the proposed method of mining activity chains.The results show that:(1).The proposed stay areas detection algorithm significantly improves accuracy by up to 35% for detecting individual stay areas from sparse mobile phone location data compared to the ST-DBSCAN and SMoT algorithms.(2).The frequency distribution of major activity chain patterns and the temporal distribution of different activity purposes in general agree well with the patterns reported by the travel survey of Shenzhen,the KL divergence of the two patterns' distribution is about 0.05.The minor differences between these two mainly come from the sampling characteristics of mobile phone location data.(3).The spatial and temporal features work together can deliver a better result in activity purpose inference than using one of them alone,but the spatial feature plays a more important role than the temporal feature.This result suggests that only using spatial feature to infer activity purposes can be a relatively reliable way.(4).The incremental of mobile phone data size can effectively improve the credibility of activity chain mining,but there is a balance between data size and data utility,researchers should choose the appropriate scale of data according to the research problem.To sum up,this study proposes an effective method to mine activity chains from large-scale mobile phone location data.The contribution of this study mainly includes three aspects.First,the study proposed a stay areas detection algorithm,which can be used to improve the effectiveness of detecting individual stay areas,and to provide reliable results for many studies based on mobile phone location data.Second,the study proposed an activity inference algorithm which can aggregate spatial and temporal features by linear interpolation.The proposed algorithm is helpful to understand the mechanism of spatial and temporal features in activity inference.Finally,the study not only lies in the methodology innovation,but also involves the thinking of taking advantage of big trajectory data with different data condition,which makes big trajectory data more practical for domain experts.The findings of this study can provide technical support for research based on activity chain,extend the application of mobile phone location data in some related fields.