Research On Overlapped Community Detection In Heterogeneous Networks

As a booming and promising research field, complex network analysis has been attracting lots of researchers' attention. Community detection, the foundation of social network analysis, can provide a new perspective to analyze the network for the researchers.The traditional technologies of community mining treat the homogeneous networks as the research subjects, and it assumes that all the nodes or links in the network belong to the same type. This assumption does make a great contribution to the study actually, but it still has a huge difference when compared with the real network in our life. In the real complex or social network, the type of nodes or links may be larger or equal to 2, therefore, we need another model to describe this phenomenon, and then the heterogeneous network was created under this requirement.Meanwhile, the communities in a real social network are usually overlapped, and it means that any node of the network can belong to more than one community at the same time.In order to discover this type of community, the techniques of overlapping community detection were proposed and have been developed very quickly. Since the overlapping community structure gives us more actual and detail description of the network, the methods which can find this type of community have a bigger practical significance. In this thesis, we study on the heterogeneous network, trying to find an efficient method to discover the overlapping community in it. The main research contents of this thesis are as follows:First of all, this thesis introduces some basic conceptions in social network analysis and has a detail analysis on those conceptions' property, such as complex network, community structure and its properties and so on. Then, we have a summary of the existed technologies for overlapping community detection. By learning and analyzing those conceptions and technologies, we determine the basic research idea.Secondly, combined with the semi-supervised learning in machine learning, we develop a mathematic model which can transform the multi-dimension network, a kind of heterogeneous networks, into a homogeneous network with a minimum information loss. The model gets a new matrix representation of the homogeneous network corresponding to the original multi-dimension network by recombining the dimensions in the original network.Since different dimension have different significance, we assign different coefficients for each dimension, so the main task of the model is to determine the coefficient vector.Thirdly, after integration of the existing overlapping community detection algorithms and information theory, we summarize a new definition of community from the perspective of information theory, and present an algorithm which can detect the overlapping community structure in complex networks after having learnt from the information dissemination theoryand the label propagation algorithm. The algorithm uses the actually accepted amount of information of the nodes as the criterion to decide whether the nodes could join the community represented by that topic.Finally, we develop some experiments and analysis on the real datasets for the proposed network transform model and overlapping community detection algorithm. For the network transform model, we use Matlab to simulate and test it, and for the algorithm, we use Java language to program and UCINet to visualize the results. After having proved the effectiveness of the model and algorithm, we do the integrated experiment on a real dataset(DBLP dataset), for this part of experiment, we firstly transform the multi-dimension network constructed by the DBLP dataset into a homogeneous network, and then apply our overlapping community detection algorithm on it. The experiments results illustrate the effectiveness and efficient of our model and algorithm when detecting the overlapping community in a heterogeneous.