Research On Key Issues Of Spreading Behavior In Microblogging Network

In recent years, with the development of Internet, many online social networks (OSNs) likemicroblogging systems emerge as a kind of novel medium platform, and provide us a new meansof communication. Many researchers from the field of computer science, physics, mathematicsand biology commit themselves to the work of OSNs analysis. However, it is very challenging inthe study about microblogging networks as the large number of participants, the frequentupdating of topics, the rapid spreading of information, and the widespread impact. In this thesis,we mainly take Sina weibo as research platform and focus on the problems of network topologyanalyzing, information diffusion modeling, key users mining, and emotion evolution modeling.This work first analyzes the influencing laws of network structure on dynamic behaviors, andthen presents some efficient strategies to control dynamic processes in microblogging network.The conclusions proposed in this thesis could be useful for the perception, tracking,early-warning, and intervening of public opinion in microblogging network.Four main contributions of the thesis are as follows:(1) We propose a united method named MixCDer by remodeling the network forcommunity detection without losing the information of link structure and user content inmicroblogging networks. First, the link direction is converted to the weight by giving highervalue to the more surprising link, while content similarity between two users is measured by theJaccard coefficient of common features and interest similarity based on LDA model. Then, abovetwo factors are uniformly converted to the edge weight of a newly generated network, which isundirected and weighted. By comparing with the partitioning results on original networks, thealgorithms of CNM, OLSOM and Infomap get better results using remodeled networks, and theaverage accuracy improves39%.(2) In order to analyze the information spreading process in microblogging network fromthe aspect of theory, we build a novel information diffusion model named MBSIT consideringboth network topology and spreading mechanism in microblogging networks. We firstly classifythe nodes to three types: Ignorant, Spreader and Teriminator, and then construct the diffusionmodel by defining the converting conditions among the three types of nodes and considering thephenomenon of repeating diffusion. Based on MBSIT, extensive experiments have been done toestimate the impacts of directedness and modulrity on the outbreak size and spreading speed of astory in many artificial networks. By analyzing the experimental results in detail, we show thatmodularity impacts the outbreak size much more than directedness. In most circumstances, theoutbreak size of a story in the network with moderate modularity could be larger than that with low or high modularity, and the coverage rate reaches85%. As a significant application, we findsome efficient strategies to control information spreading in real-world microblogging networks,and a better outcome will be attained to improve or restrain the outbreak size by adding orremoving the edges between modules, rather than the edges with large betweenness.(3) For the problem of influence ranking among the users related with a specific topic inmicroblogging network, we design and implement a topic-sensitive influence ranking algorithmnamed WTSIRank. WTSIRank first uses a node-weighted graph to illustrate the social network.Weight value is mainly determined by the content and the number of the published posts. Wealso design and impletement a method to calculate the node weight based on correlationmeasuring and originality distinguishing. The transition probability between individuals iscalculated using the correlation coefficient between the published posts and the concerned topic,and use the node weight to fix the influence score of each individual. With two real-world datasets from Twitter, we compare the algorithm of WTSIRank with DegreeRank, PageRank,LeaderRank and TwitterRank. From the comparing results, we find that the highest correlationcoefficient between WTSIRank and other algorithms is0.71, and WTSIRank outperforms allother methods and could exploit the key individuals with more accuracy.(4) We design and implement an emotion classifier named MBECer based on Bayes theory,build a dynamic evolution model of collective emotions named EDEM, and present an efficientstrategy to guide the collective emotions. MBECer is designed by constructing a corpus andcalculating the probability of each word in different emotions. Considering the features ofChinese, two optimization strategies based on Entropy and Salience filtering are presented toimprove the performance of MBECer. From the test results, we know that the accuracy ofMBECer without optimization is59%, and the optimized accuracy is83%, which is better thanthe classifying result on Twitter (81%). By analyzing the collective emotions in our samplenetworks in detail, we get some interesting findings, including a phenomenon of emotionspreading between friends. Furthermore, we find that the number of friends has strongcorrelation with individual emotion. Based on those useful findings, we present a dynamicevolution model of collective emotions, in which both self-evolving process andmutual-evolving process are considered simultaneously. As an application of our emotiondynamic model, we design an efficient strategy to control the collective emotions of the wholenetwork by selecting seed users according to k-core rather than degree, and the guiding resultaveragely improves10%.