Community Detection Based On Topology Potential And Spectral Clustering

Community detection is a hot topic in research of complex networks. There are many kinds of community detection algorithms. Spectral clustering is one kind of typical. Compared with other traditional methods, it has simple mapping, rigorous mathematical logic, and is applicable to any shape data sets and so on. However, there are also deficiencies. Firstly, traditional spectral clustering must be manually determined the number of communities. Spectral clustering based on normal matrix uses the trapezoidal structure of community of the spectral gap to identify the number of communities, but when the network community structure is not obvious, the number cannot be judged correctly by spectral gap of normal matrix. Secondly, mostly to construct the mapping matrix of spectral clustering is used by node degree, which cannot reflect relationship between node and the whole network, which means it cannot contain more structural information, leading to devious results of eigenvalues and corresponding eigenvectors, affecting the community division results..This paper presents the number of complex network community detection method based on topological potential. This method calculates topology of all nodes. Our method derives local maximum potential value node from wave crest and trough, obvious network and bump is not obvious network by convex-concave parameter. Then through the appropriate method to judge whether merging in these extreme points or not. Finally get the number of community of whole network. Through the experiment by real complex networks and artificial generation network, shows that this method has a higher accuracy on number of community detection.This paper presents the community detection based on topology potential and spectral clustering algorithm. It improves spectral clustering based on the normal matrix by using topology, through obtaining the topology matrix, which can reflect the close relationship between the node and global network, make matrix contain more structure information. Meanwhile, get local extreme points as the initial center points in K-means clustering algorithm, which reduces the number of iterations, improves the stability and accuracy of spectral clustering algorithm.