Analyzing And Modeling Of Popularity Of Social Media Content

Social media has been experiencing rapid developments in recent years, and is nowamong the mainstream in the cyberspace. According to the Alexa statistics, out of the topten sites with the heaviest traffic around the world, five are social media sites. Theunprecedented development of social media has been promoting abundant new researchfields. In2009, Science magazine published a landmark work entitled "ComputationalSocial Science", which marks the cross-integration of computing science and social scienceis becoming one of the hottest global research forefronts. The study of distribution anddynamics of popularity of social media content is one of the most important research areasand is now attracting more and more endeavor from research community. The explorationof distribution and dynamics of social media content popularity not only deepens theunderstanding of the laws of humankind’s collective behavior, but also strengthens thetheoretical foundation for understanding and enhancing performances of critical servicessuch as pre-fetching and caching system, P2P networks, search engines andrecommendation systems. This dissertation conducts an in-depth study on issues ofanalyzing and modeling popularity of social media content, including characterizing ofpopularity distribution, modeling of popularity propagation mechanism, improvingpre-fetching and caching policy based on the characteristics of popularity distribution, andproposing effective method to improve popularity of social media content.First, several characteristics of the distribution of social media content popularity aswell as the impact of each source on the distribution are analyzed. With a huge scale,User-Generated Content exhibits distinctive features like highly dynamic and highlyfragmented, may make the traditional popularity distribution model and prediction methodsinvalid. We differentiate our work from previous studies in the following two aspects.Firstly, we investigate the overall distribution characteristics of multi-source popularityfrom both global and local perspectives. Secondly, we distinguish the popularity from eachsingle source, which has never been done before, and thus enable us to investigate theimpact of each source on the popularity distribution. Our investigation shows that search engines and recommendation systems are the two main sources of popularity of socialmedia content. Moreover, search engines tend to raise skewness of popularity distribution,while recommendation systems help reduce the skewnewss.Secondly, a novel multimedia content pre-fetching and caching method based onClustered User Behavior Model (CUBM) is proposed. Even though user generated videosharing sites are tremendously popular, the experience of the user watching videos is oftenunsatisfactory. Delays due to buffering before and during a video playback at a client arequite common. We motivate the need for pre-fetching and caching by performing aPlanetLab based measurement demonstrating that video playbacks on YouTube are oftenunsatisfactory. On this basis, we then propose a novel multimedia content pre-fetching andcaching method based on Clustered User Behavior Model (CUBM). The proposed methodis capable of classifying users with similar behavior patterns into clusters and thenestablishing a respective Markov chain to represent browsing patterns for each cluster ofusers. As a result, the method is able to overcome the shortcomings of the traditional singleMarkov chain method that fails to express the differences between individuals, as well as toseize the fact that active users tend to browse more content, thus leading to a considerableimprovement in pre-fetching accuracy.Thirdly, a random walk based popularity propagation model, RWPPM in abbreviation,is proposed. Our goal is to study how one item’s popularity impacts popularity of otheritems via networked links. Usually, the click through rate from one item to another item isnot difficult to obtain. However, this probability is not sufficient for our goal as it onlycaptures the relationship between an item and another item that directly linked. Therefore,we propose a popularity propagation model, which is able to derive the influence of oneitem to other items, even though there are no direct links between them. After analyzing theconvergence conditions and validating the accuracy of the model, we then take the YouTubevideo network as the example to study the mutual influence between the items in anetworked background. The model is applicable in analyzing a wide range of relationshiptypes, such as online social networks, blog networks, as well as academic citation networks.Finally, we propose a keyword suggestion algorithm for suggesting keywords that arerelevant to a topic given by a user and have high potential to improve popularity of socialmedia content. We analyze the role of titles and tags of multimedia items in helping search engines and recommendation systems retrieve multimedia items. We then demonstrate thatmultimedia items with similar topics tend to form a cluster in the referrer graph induced bythe recommendation system and exploit this property to obtain keywords relevant to a topic.On this basis, we then propose a keywords recommendation algorithm that follows theideology of "Keywords-Topics-Keywords" and ranks keywords based on both theirrelevance and potential to attract views. Our case study experiment demonstrates theeffectiveness of the keyword suggestion algorithm in increasing item popularity.