Research Of Clustering Attributed Graphs Based On Multi-feature Fusion

With the development of internet, more and more requirements of social communication are emerging. As a result, a huge of datasets from social networks are produced. In general, these datasets contain the relationships among users and the basic information of everyone. Existing graph clustering studies for mining social networks almost only focus topological relationship of nodes, but seldom take consideration of node attributes. Therefore, it has lots of research and application meanings for mining graphs with node attributes.In this paper, we name the graphs with node attributes as attributed graphs. It can be categorized into two parts, one is single attributed graph and another is multiple attributed graph. Single attributed graph is a graph with a single node attribute view, multiple attributed graph is a graph with multiple node attribute views. It is the key research theme to how to integrate topological relationships and node attributes in reasonable and effective way.To solve the single attributed graph clustering problem, we propose a joint weighted nonnegative matrix factorization(JWNMF) clustering method. The method integrates topological and attributes into a uniformed objective function, and also takes the weights into account. By solving the object function, we can obtain a factorized matrix with both topological information and attribute information. And we only need to make clustering on this matrix, then the single attributed graph clustering problem can be solved. Meanwhile, we also prove the convergence of this method. Finally, the results of experiments show that our JWNMF can outperform some existing single attributed graph clustering methods.To solve the multiple attributed graph clustering problem, we propose a multiple joint weighted nonnegative matrix factorization(MJWNMF) clustering method. Since node attributes have multi-view property, we use the method for dealing with item-user relationships which proposed in heterogeneous collaborative filtering(Hete-CF), to integrate the node attributes from diverse views. Then extend JWNMF to integrate topological information and multiple node attribute views into a uniformed objective function, and as what we do in JWNMF, make clustering on the factorized matrix. Finally, the results of experiments show that our method more useful on those multi-view datasets with topological relationships.