Exploring network structures in feature space

We propose a multi-phase approach to explore network structures. In this method, structure analysis is not carried out on the observed network directly. Instead, certain similarity measures of the nodes are derived from the network firstly, which are then projected onto an appropriate lower-dimensional feature space. The clustering structure can be defined in the feature space, and analyzed by conventional clustering algorithms. The classified data are finally mapped back to the original network space if necessary to complete the analysis of network structures. By mapping onto the feature space, some difficulties due to the diversity of micro-structures and scale of the network can be circumvented. This makes it possible for the proposed method to deal with more general structures such as detecting groups in a random background, as well as identifying usual community structures in networks.