Detecting Human Mobility Periodical Patterns From Social Media Data

Economic development has brought heavy urbanization, which leads more and more people to move into large cities. Additionally, the development of public trans-portation has also enabled people to move from one place to another more easily and quickly. Consequently, human mobility in urban areas has become more frequent and complex. Although these human mobility patterns are seemingly random and hard to predict, the majority of people have some latent regularity in common. Dis-covering such latent human mobility patterns is very meaningful to understanding or solving related problems, i.e., human migration prediction, the spread of dis-eases, optimization of immunization methodology, human contact patterns, urban planning, and anomaly detection. Thus, understanding human mobility patterns is extremely important.Human mobility studies go back to 1885, when E.G. Ravenstein published The Laws of Migration. In this book, he discovered a few very important and fundamen-tal laws that govern human migration:1) Most migrations cover short distances; 2) people tend to move into developed areas, a process called absorption; 3) there is an inverse phenomenon of absorption, called dispersion, and 4) economic factors are the main cause of migration. Many studies have been done to verify and formalize these observations via quantitative research.However, understanding real-world individual mobility patterns is a very diffi-cult task because of the lack of individual location data on a large population scale. Traditional mainstream data sources, i.e. survey-based data and census data, is high cost, coarse both in temporal and geographical aspects. Thus, it is hard to analyze human mobility, which is temporally and spatially dynamic, with these static data. With the advent of location-based social networking services, millions of users come to post and share their status with their location information on these social net-works. As a result, a huge amount of spatiotemporal data is available and such data offer new opportunities to study human mobility patterns at an individual level by the characteristics of the data, namely, its fine location granularity and the large population scale of each individual.However, these spatiotemporal data are not complete for a number of reasons. One of the biggest problems is that this new data is very sparse since users do not necessarily post their status or share locations on such location-based social networks every time they move. In this thesis, to address this problem, we assume that there is daily periodicity in human activities, which has been verified by transportation usage and location visitations in related research. Based on this assumption, the spatiotemporal data collected over a long period of time can be viewed as one day of activity by putting all the data into a single day period, ignoring the year and date of the timestamp of the data.This thesis includes three main parts:1) We choose the largest social blogging service, Twitter, as a data source, which provides us with access to its streaming Twitter data with location and time information. Based on the aforementioned assumption, we perform an analysis of individual use of Twitter and user movements in Tokyo, the largest city in the world. We observe that there is a difference of Twitter use between weekdays and weekends, which represents a difference of daily activities during these periods. Next, we identify users’frequently visited locations based on those that are dominant in their tweeting location histories and are assumed to represent users’lifestyle and daily routines. Our simple model fitted for each hour discovers most users have two or three frequently visited locations as reported in related works. After understanding users’long-stay locations, we analyze temporal patterns by utilizing their detected frequently visited locations. (Lastly, we integrate spatial and temporal characteristics to build our spatio-temporal model of users’tweets patterns that not only captures individual trajectories but also classifies the status of each user at a given time.)2) We developed a trajectory visualization application, which can plot geograph-ical points where each user posted their status and show their captured trajectory patterns on the map with basic statistical information about the users. This web application uses Ruby on Rails for the web framework, MySQL for the database, Thin for the web server, and Nginx for the http server.3) We consider a social network-based recommendation system that provides the advertisement strategy based on the captured human mobility patterns of each user. Specifically, the recommendation system takes the location and state of the user into consideration, which enables to recommend more suitable and related advertisements.