Analyse et application de la diffusion d'information dans les microblogs. (The analysis and applications of information diffusion in microblogs)

Microblog service (such as Twitter and Sina Weibo) have become an important platform for Internet content sharing. As the information in Microblog are widely used in public opinion mining, viral marketing and political campaigns, understanding how information diffuses over Microblogs, and explaining the process through which some tweets become popular, are important.The analysis of the information diffusion in Microblogs involves the data collection from Microblog, the modeling on information spreading and using the resulting models. Dealing with the huge amount of data flowing through microblogs is by itself a challenge. Designing an efficient and unbiased sampling algorithm for Microblog is therefore essential. Besides, the retweeting process in Microblog is complex because of the ephemerality of information, the topology of Microblog network and the particular features (such as number of followers) of publisher and retweeters.Two traditional models have been used for information diffusion : Independent Cascades and Linear Threshold models. However no one of them can describe completely the retweeting process in Microblog accurately. The analysis and design of new models to characterize the information diffusion in Microblog is therefore necessary. Moreover, a comprehensive description of the correlation between the information diffusion in Microblog and the searching trends of keywords on search engines is lacking although some work has been found some preliminary relationships.This work presnets a complete analysis of information diffusion in Microblog from. The contributions and innovations of this thesis are as follows:1)There are two popular unbiased Online Social Network (OSN) sampling algorithms,Metropolis-Hastings Random Walk (MHRW) and Unbiased Sampling for Directed Social Graph (USDSG) method. However they are both likely to yield considerable self-sampling probabilities when applied to Microblogs where there is local. To solve this problem, I have modelled the process of OSN sampling as a Markov process and have deduced the sufficient and necessary conditions of unbiased sampling. Based on this unbiased conditions, I proposed an efficient and unbiased sampling algorithms, Unbiased Sampling method with Dummy Edges (USDE), which reduces strongly the self-sampling probabilities of MHRW. The experimental evaluation demonstrate thats the average node degree of samples of MHRW and USDSG is 2 - 4 times as high as the ground truth while USDE can provide the approximation of ground truth when the sampling repetitions are removed. Moreover the average sampling time per node in USDE is only a half of MHRW and USDSG one.2)A second contribution targets the shortages of Independent Cascades (IC) and Linear Threshold (LT) models in characterizing the retweeting process in Microblogs. I achieve this by introducing a Galton Watson with Killing (GWK) model which considers all the three important factors including the ephemerality of information, the topology of network and the features of publisher and retweeters accurately. We have validated the applicability of the of GWK model over two datasets from Sina Weibo and Twitter and showed that GWK model can fit 82% of information receivers and 90% of the maximum numbers of hops in the real retweeting process. Besides, the GWK model is useful for revealing the endogenous and exogenous factors which affect the popularity of tweets.3) Motivated by the correlation between popularity and trendiness of topicsin Microblog and search trends, I have developed an economic analysis of the market involving a third-party ad broker, which is a popular market in current SEM, and finds that the adwords augmenting strategy with the trending and popular topics in Twitter enables the broker to achieve, on average, four folds larger return on investment than with a non-augmented strategy, while still maintaining the same level of risk.