ジオタグつきtweetの時空間解析に基づいた地域特性抽出に関する研究

抄録

 This paper clarifies how people get contact with information space in their daily life through tweet data, and identify characteriistics of each area from the viewpoint of relationship between information space and substance space. Recently, flow between these two spaces have become activated through technical innovation, and Twitter represents the relationship. This paper defines temporal-spatial attributes of tweets and describes regional characteristics by utilizing the attributes. In particular, we suggest a method to identify user's residenatial/working areas based on their tweeting pattern, and moreover, visualize the potential distributions of each area.
 This paper can be divided into two parts. One is identification of tweets' attributes, and the other is visualization of regional characteristics. There are various studies on identification of tweets' attributes especially identification of user's residential area by analyzing tweet data, but most of them need to take advantage of some locar words/data, therefore are lacking in universalities. On the other hand, this paper enables comprehensive estimation by only using tweet's location and time data. Moreover, while most of studies on visualization of regional characteristics treat tweet data just as substitute for other geographic data (e.g. person-trip data, GPS data), this paper utilizes characterisitcs of Twitter and introduce new attributes of tweets.
 In this paper, we acquired 12,553,361 tweets between March 26th - April 28th. 2015 covering all parts of Japan. By calculating these data by user, we analyze from each viewpoint of stagnation, continuousness of tweeting, and exogeneity. First, we define tweet distance as a distance between locations of two successive tweets by the same user, and tweet span as an interspace between times of two successive tweets by the same user. By examining these two values, we can evaluate tweets' stagnation and contiuousness of tweeting. Therefore, short tweet distance means the user stagnate at one spot and short tweet span indicates that the user post tweets in a row. Next, we estimate tweets' exogeneity by identifying users' main tweeting points. Through spatial clustering, we acquire two main clusters for each user and define their center points as main tweeting points. Then, we categorize the two points into daytime/nighttime main tweeting points by modal tweeting time of each main cluster. Based on the distribution of daytime/nighttime main tweeting points, it can be said that daytime main tweeting points represent users' working area (or school) and nightime one do users' residential area. Actually, average distance between daytime/nighttime main tweeting points is ~30km, which is roughly equivalent to average distance between working space and residential area in Tokyo.
 In chapter 5, we visualize the regional characteristics by utilizing the three tweets' attributes introduced in the previous chapter. In concrete terms, we select two areas in Tokyo (Harajuku and Akihabara) and visualize the potential distribution for each. First, we apply Kernel density estimation to the areas and observe the tweet-density. Moreover, by levaraging Kriging metohd, we enable to visualize stagnation potential, continuous-tweeting potential, and exogeneity potential. From the viewpoint of projection from information space, we can not only reacknowledge the images of each area, but also discover some new regional characteristics which have been veiled in the substance space.
 This paper realizes the method to evaluate the substance space from the viewpoint of information space. It is possible to utilize this method and knowledge for various purposes from urban planning to marketing.