Attractive Force Field Modeling Method Of Complex Networks And Its Applications

Complex networks can be used to disclose several features of complex system-s, such as their structures, functions and the interplay between them. Recently more and more researchers are interested in the studies of complex networks, ranging from physics, biology and economic to sociology. However, most of the existing studies rarely build model and solve the related problems from the standpoint of physical sci-ence. Therefore, in this thesis we treat vertices in networks as particles of spaces, and each vertex would produce some kind of force on all the other vertices. And then sev-eral important issues in the current studies of complex networks can be solved from another point of view by building the virtual attractive field formed by components of the corresponding network.In this thesis, firstly, we assign several physical attributes to vertices in networks from the perspective of physical mechanics. Complex networks are regarded as me-chanical systems in virtual spaces and the attractive field model is built by calculating the virtual forces among nodes and the related changes of inner relationships in networks caused by attractive forces through a series of theories and methods. The attractive field model shows the natural attributes of data and could be applied to deal with problems of networks in data mining and other fields by choosing different parameters.Secondly, most existing community detecting problems are optimization or heuris-tic methods, resulting in a low convergence speed and non-ideal division accuracy of communities. Therefore, a novel definition of community structure based on the at-tractive field model of complex networks is proposed and its intrinsic characteristics are analyzed. Furthermore, we put forward an iterative algorithm NC-DF (Network Communities in Data Field) to analyze and detect community structures. The proposed method makes full use of the clustering property represented by the attractive force between nodes, and it can be conducted totally adaptively without any preset param-eters. Besides, it’s able to converge fast and acquire a stable division result, showing preferable performances on both synthetic and real-world networks.Finally, most algorithms on link prediction just exploit the individual information of common neighbors and lack the consideration of mutual interactions among com-mon neighbors. To solve this defect, we propose a new similarity index based on the attractive strength of arbitrary pair of nodes in the attractive field model, i.e., the at-tractive density of the cluster formed by common neighbors. The proposed index uses the attractive density to measure the similarity between each pair of nodes, which not only considers the individual information of common neighbors, but also exploits the interactions among common neighbors. It improves the prediction accuracy while guar-anteeing the time complexity, especially on networks with lower clustering coefficients.