An Improved Spectral Clustering Algorithm for Community Discovery

For discovering communities in social network, an improved spectral clustering method is presented in this paper. To make full use of the network feature, the core members are used in this method for mining communities. This goal has been achieved through the Page Rank method, which is common in directed graphs, for the reason that an undirected graph can be treated as the special case of the corresponding directed one. Following that, they can be used for initialization in the spectral clustering to avoid the sensitivity to the initial centroids. Applied to four datasets, the improved method turns out to be better than the traditional spectral clustering methods, whether in time or in accuracy aspect.