Research On Similarity Representation And Its Applications In Complex Networks

The notion of ‘complex networks' came to use in late 1990 s and lots of computer scientists,biologists,sociologists,physicists,and mathematicians started to intensively study diverse real-world networks and their models.The field of complex networks is currently a very hot and attractive research area.Link prediction and community detection are two hot topics in this area.One of the methods for judging whether two nodes have connections or belong to the same community in the network is to calculate the similarity.This method has low computational complexity and it's more suitable for large-scale networks.There are many similarity index proposed until now,but most of them only consider degree of the node and its common neighbors which leads to insufficient prediction accuracy,and the others are not suitable for complex networks because of their computational complexity.Based on Deepwalk and clustering algorithm,the similarity representation is studied and applied in this paper.The main contents are as follows:(1)Because of the high computational complexity of global similarity and the limited prediction accuracy of local similarity in traditional similarity representation,a new similarity index,Deep Affinity(DA),is proposed in this paper.It combines the traditional similarity index with the distance index obtained by Deepwalk and the concept of clustering.The coordinates of nodes and Euclidean distances between nodes are obtained by Deepwalk and K-means is used to get the cluster number of each node.Finally,the similarity of the two nodes is obtained by coordinating the influence degree by the two parameters ? and ?.After conducting experiments on different network datasets,the results show that DA-based link prediction algorithm has greatly improved the prediction accuracy,especially for large-scale network datasets.(2)Many proposed community detection algorithms can only achieve better results on small networks,but not on large networks.This paper proposed a new local community discovery algorithm based on DA index,called CBB,finds communities by identifying borderlines between them using boundary nodes.Our method performs label propagation for community detection,where nodes decide their labels based on the largest ‘‘benefit score'' exhibited by their immediate neighbors.The proposed algorithm has a local approach and focuses only on boundary nodes during iterationsof label propagation,which eliminates unnecessary steps and shortens the overall execution time.The algorithm has a distributed nature and can be used on large networks in a parallel fashion.Experiments on real-world network datasets and artificially generated LFR datasets show that in the quality of identified communities,CBS performs better than CBB proposed by Tasgin M etc.in 2018.