Information Diffusion Models In Micro-blogging Networks Based On Hidden Markov Theory And Conditional Random Fields

With the rapid development of social networks, we enter into the era of big data as theexplosive growth of information. In order to better find the potential value of social networks,many scholars have conducted research on various aspects of them. And how to takeadvantage of social networking information and effectively control or guide its proliferation?How to deeply understand the diffusion mechanism of social networks? How to effectivelypredict user behavior in social networks? A pair of well-established research direction is toestablish accurate information diffusion model.Micro-blog, as a new social networking platform, has commonalities of traditional socialnetworks and its own personality. So far, studys on information diffusion model inmicro-blogging networks which consider the information content, micro-blogging users aswell as the network structure are a handful. Furthermore, studys on the multi-informationdiffusion considering both "competitition" and "collaboration" among information, and alsoare based on statistical probability is rare. In view of this, we propose an information diffusionmodel based on the hidden Markov theory (IDMBHMT) and a multi-information diffusionmodel based on conditional random fields (MIDMBCRF) in micro-blogging networks.Firstly, through comprehensive studys of the characteristic and influencing factors aboutthe information diffusion in micro-blogging networks, hidden Markov theory, ConditionalRandom Fields (CRFs), and methods related to the usage of characteristic functions in thispaper, including automatic Chinese text categorization, user similarity measure andinformation interaction quantitative method, we set up a information diffusion model based onhidden Markov theory (IDMBHMT) and a multi-information diffusion model based on CRFs(MIDMBCRF) in micro-blogging networks. Secondly, we used the METIS tool to partitionthe user relationship network in micro-blog, and set up the two models according to thesub-network to improve the performance of them. Furthermore, we applied our models topredict the user’s reposting behavior using the junction tree algorithm. Finally, we did thesimulation experiments according to the dataset crawled from Sina micro-blogging using itsAPI(Application Programming Interface). Through the experiments, we analyzed the factors that affect performance about thesetwo models: graph partitioning technique do improve the performance of the two models, andwhen the sub-network size equals to48, the two models reach their peak performance; inaddition, the reposting probability of MIDMBCRF on average43%comes from “interactionsbetween information pieces”. Besides, when the network size is in the case of120,240and400, comparing IDMBHMT and MIDMBCRF with Jiang’s retweet prediction model based onlogistic regression(RPMBLR) and Zheng lei’s multi-information diffusion model based onLinear-Threshold (MIDMBLT) respectively, we found the descending order for the fourmodels according to their performance is: MIDMBCRF, IDMBHMT, RPMBLR, MIDMBLT.In summary, we can see the two information diffusion models constructed in this papercan not only be applied to predict the users’ behavior and guide public opinion, but also havegreat merits in studies on other associated disciplines.