Research On Information Diffusion In Online Social Networks

Recently,with the development of Internet technology,the online social network plays an important role in people's daily life.The online social network is the representation of real world network which consists of the individuals and the relationships among individuals.It means online social network is the reflection and extension of real world.The information about the event which happens in the real world will exist in online social network.And it will propagate and evolve with the interactions among the users in online social network.And what's more,it will react on the real world and influence people's behaviors and actions.The online social network narrows the gap between the individuals,accelerates the propagation speed and fertilizes the relationships among individuals.So the online social network becomes one of the research highlights in the field of sociology,communication,computer science,system science,and etc.The analysis of information diffusion in online social network is related to natural language processing,data mining,machine learning,communication,sociology,and etc.In online social network,there are many types of data,complex relationships and fast propagating information.Those characteristics make it different for information diffusion analysis in online social network from the traditional analysis.It brings new challenges as well as new opportunities.This paper aims at influence maximization,real-time personalized search,text classification and information diffusion learning based on the existing related work.The main achievements are as follows:1.In terms of influence maximization problem,this paper proposes the influence efficiency maximization problem,analyses the complexity of that problem and design a reverse efficiency sampling algorithm to solve that problem.Traditional influence maximization problem only take the final influence into consideration,neglecting the time stamp that node is activated,which is called propagation time delay.While in practical application,it is meaningful to activate nodes as soon as possible.To solve the problem,we take propagation time delay into account,define the influence efficiency function and propose the influence efficiency maximization problem.After that,we prove that problem is NP-hard under independent cascade model.And what's more,the computation of influence efficiency is #P-hard under independent cascade model.We also prove that the influence efficiency function is submodular.At last,we design a reverse efficiency sampling algorithm to solve the problem.2.In the terms of real-time personalized searching,this paper propose a framework for real-time personalized searching which integrates semantic extension and quality model.In traditional information retrieval model,the system returns the re-ranked results based on user input.While in online social network,the information generates in high speed and emerges as data stream.Moreover,it is proper to push information automatically according to user's intention.To solve those problems,we propose a framework for real-time personalized searching that integrates semantic extension and quality model.We adopt a user model based on boolean logic keyword filter.It constructs based on external knowledge base,leverage the query extension to re-rank the results according to the relevance.Furthermore,we utilize a text quality model based on logistic regression.It evaluates text's quality based on the social attributes,making the retrieval text meaningful.3.In the terms of text classification,this paper proposes a text classification algorithm based on text summarization and convolutional neural network.In online social network,there are enormous text data and a large variety of topics.So traditional bag-of-word model confronts the problem of dimension explosion.Besides,there is context-dependent in the text.While traditional methods make no use of that property.To solve that problem,we propose a text classification algorithm integrating word vector model and convolutional neural network.The algorithm fixes the dimensionality and keep the local feature of the context.It includes text summary extraction,text vectorization and convolutional neural network.Finally,the experiments verify the effectiveness of the algorithm.4.In the terms of diffusion model learning,this paper proposes a diffusion model based on probabilistic reading.In practice,an event consists of many information.So we need to take multiple network convergence into consideration.And furthermore,the spam users in online social network bring noise into the evaluation of influence.So we need to construct a spam user filter to get rid of the noise.Moreover,whether the user receives the information or not is probabilistic when the information is pushed to user.So we need to model user's reading probability.To solve those problems,we propose a diffusion model based on probabilistic reading,which considers those problems comprehensively.After training,the model can reflect the actual influence of an event.In a conclusion,this paper aims at influence maximization problem,real-time personalized searching,text classification and diffusion model learning in online social network.We carry out experiments on real world dataset,and the experimental results show the effectiveness of our methods.The methods for information diffusion analysis in online social network are practical.