A Study on Patterns of Information Cascades in Microblogs based on Distributions of Users'Influence and Posting Behaviors

As online social networks become extremely popular in these days, people communicate and exchange information for various purposes. We realize that different activities tend to have different ways information spread on the network. Knowing patterns of information cascade would help organizations to examine behaviors of public relation campaigns. In this thesis, we perform a research on Twitter's user network to understand patterns of information cascade and behaviors of participating users in various topics. We verify whether different topics really have different cascade patterns or not by exploring four measures, which are cascade ratio, tweet ratio, time interval, and exposure curve. We conduct experiments on a real Twitter dataset. We consider Twitter hashtags as representatives of topics and obtain six major topics, which are earthquake, media, politics, entertainment, sports, and idiom. We firstly study the pattern of hashtag cascades in each topic by using statistical approach, then further investigate the relationship between cascade patterns and topics by using clustering algorithm, and lastly verify the effectiveness of each measure due to the clustering results. Our experiments show that hashtags in different topics have different cascade patterns in term of cascade ratio, tweet ratio, time interval, and exposure curve. For example, the earthquake topic has low cascade ratio, low tweet ratio, short lifespan, and high persistence, while the political topic has high cascade ratio and high persistence. However, some hashtags even in the same topic have different cascade patterns. For instance, the earthquake hashtags can be divided into the hashtags directly related to the Great East Japan Earthquake, the media-related hashtags, and the political-related hashtags or the hashtags about the nuclear power plant. We discover that such kind of hidden relationship between topics can be surprisingly revealed by using only four measures rather than considering tweet contents. Finally, among four measures we explored, our results also showed that cascade ratio and time interval are the most effective measures to distinguish cascade patterns in different topics, while tweet ratio and exposure curve from the related work are not effective as we expected.報告番号: ; 学位授与年月日: 2012-03-22 ; 学位の種別: 修士 ; 学位の種類: 修士(情報理工学) ; 学位記番号: ; 研究科・専攻: 情報理工学系研究科電子情報学専