Research On Information Diffusion Predictive Models In Online Social Network

Nowadays, online social network increasingly plays a significant role in fields of politics, ecnomy, culture and commercial. It is urgently important for either research commumity or industry to study problems related to online social network. Information dissufion is the most important one among the series of unresolved issues. On the one hand, we can more effectively utilize online social network to spreading information by understanding the diffusive regularities, on the other hand, we can avoid impact from adverse message such as fraud and rumor so that decrease the possible threat to public society.This dessertation gives out a detailed discussion on information diffusion over online social networks and mainly concducts a plenty of study on predictive diffusive models. Be different with a couple of similar research works, the dessertation focuses on a spatial-temporal information diffusion problem from not only spatial but also temporal dimension. The spatial-temporal diffusion problem is that for a given information m initiated from a user s called source, after a time period of t, what is the density p of influenced users at distance x from the source s. With sufficient experimental analysis on real Digg dataset, our study work discovers the diffusion characteristics and builds two mathematical models to describe the spatial-temporal information diffusion, and use the models to predict the diffusion process in online social networks.Firstly, the dessertation introduces the spatial-temporal information diffusion problem and gives out the related specific description, studies the spatial-temporal characteristics of diffusion process in real Digg dataset given a definition of distance which quantitatively measures the distance between two users in online social network graph. By conducting a series of experiments and analysis, we reveal some regularities of information diffusion from both spatial and temporal perspective. Then we build a kind of Fick-Logistic information diffusion model based on Fick's diffusion law and Logistic growth model to describe the information diffusion process according to the empirical study results above. We also evaluate the performance of the model by predicting a few most representative news stories in real Digg dataset and the results indicate that the model can achieve a high accuracy of prediction for these cases of news.Secondly, the dessertation furtherly studies the spatial- temporal characteristics in Digg dataset and design another diffusion model called inear diffusive model based on partial differential equation. The linear diffusive model takes into account influence of both spatial - distribution of user and temporal information interest decay so that the model can better describe the information diffusion process in online social network. The experiments results show that linear diffusive model have a higher prediction accuracy compared with Fick-Logistic information diffusion model.In addition, the dessertation explore an effect valuable spatial-temoral diffusion issue in online social networks called multi-source information diffusion. Base on empricial study, we introduce a new issue called superposition effect of multi-source information diffusion. Specifically, we first define the distance between two users for scenario of multi-source information diffusion, and then propose an approximate multi-source selecting algorithm to extract satisfied information cases from real Digg dataset. We find that the spatial-temporal characteristics of these multi-source information cases are very similar to the observation in study of single source information diffusion, which leads us to utilize the linear diffusive model to predict the superposition effect of multi-source information. All results indicate that proposed algorithm is simple but effective and feasible, and further prove the linear diffusive model has good capability of predicting information diffusion in online social networks.Finally, the dessertation concludes several main points and contribution of the work in this study and outlines the future work.