Influence Maximization Of Social Networks Based On Online Machine Learning

As the Internet becomes more and more developed,more and more users express opinions and spread content on social software.Users will be influenced or influenced by others.How to find the influence from social networks to maximize the collection of users to achieve Maximize the spread of information in social networks to maximize its impact.However,the current social influence maximization problem does not take into account the dynamic changes in the influence between users in the graph network.This article uses online learning to explore the characteristics of unknown parameters,and for dynamic networks,explore the parameters and obtain the seeds for maximizing influence.User collection.The main work of the thesis includes:It introduces the online influence maximization problem and the multi-arms bandit problem and algorithm used in detail,and analyzes how the combined multi-arms bandit algorithm is applied to the online influence maximization problem.In the online influence maximization problem,when the influence of the business on the users in the social network is unknown,the corresponding strategy is adopted to obtain the set of seed users within a limited round,so as to maximize the influence of the communication target.Finally,experiments based on Flickr and NetHEPT datasets and compare the performance of common algorithms.This thesis takes into account the dynamics and the diversity of changes in the influence of users in social networks.The problem of strategy selection under dynamic networks where the influence of users in social networks changes over time is based on non-stationary bandit.The research of D-LinUCB(Discounted Linear UCB),combined with the problem of online influence maximization,proposed a dynamic online influence maximization algorithm DIM(Dynamic-online Influence Maximization).The DIM algorithm detects environmental mutations in a dynamic environment through historical averages and adopts The discount strategy updates the node parameters,and obtains the best set of users in the changed social network through the method of online combination learning.This thesis evaluates the performance of the DIM algorithm based on experimental simulations on the Flickr and NetHEPT social network data sets,and proves the superiority of the DIM algorithm by comparing the cumulative benefits of different online influence maximization algorithms.In the traditional online influence maximization problem,exploring multiple communication goals to find user sets,high spatial complexity and long learning time.This thesis uses clustering and similarity migration to propose a category-based CIM(Category-online Influence Maximization)algorithm.The CIM algorithm uses the characteristic information of different propagation targets to use k-mean clustering and explore,and according to the differences between different products The similarity is entered into the network in advance after the product has been explored several times.The parameter learned by the product before the migration solves the problem of the cold start of network parameters and accelerates the convergence of online influence maximization.This thesis uses Flickr and NetHEPT social network real data sets to experiment and simulate the online influence maximization of multiple communication targets.By comparing the learning speed and memory overhead of traditional schemes,it proves the superiority of the CIM algorithm.