Study Of Influence Maximization In Social Networks Based On Autoencoder

With the rapid development of mobile Internet and big data technology,social networks have become an important channel for people to share their information.Influence maximization is the problem to find a small number of optimal individuals which maximizes the spread of influence under a cascade model in a social network.In addition,the influence maximization problem has been widely applied in advertising,public opinion monitoring,disease prevention and other fields.It has become a hot research area in network science.Finding the mo st influential individuals has been proved to be NP-hard.Among the existing algorithms,greedy algorithms usually achieve higher accuracy,but they are with high computational complexity.The heuristic algorithm can be run quickly.However,It is easily trapped in local optimum.Moreover,the existing algorithms have limitations for learning influence features of individuals.Because deep learning has the powerful feature representation abilities,this thesis utilizes deep learning to deal with the influence maximizing problem.In this thesis,we analyze information diffusion characteristics in the network,and we construct the approximate matrix which not only captures the social network structure,but also reflects the influence characteristics of individuals.An autoencoder based learning model is used to exploit the influential representation by reconstructing the approximate matrix.The learnt representation can well represent the influence of individuals in the network and we can use it to select seed nodes,instead of choosing seed nodes greedily.Experimental results on several real-world social networks show that our proposed algorithm can find the most influential set of seed nodes.In order to reduce the time complexity of algorithm,this thesis proposes a new method to measure the importance of nodes and preselect nodes in the network,which is benefit for reducing the searching space while preserving the maximum influence spread.In addition,this thesis further adds random noise to the approximate matrix,and learns the more expressive and robust node influence representation through the denoising autoencoder.The experimental results show that the node pre-selection strategy can reduce the selection range of nodes by about 40%,and influence diffusion results will not be affected by the initial seeds.