Detecting hierarchical structure of community members in social networks

Current methods often predefine fixed roles of members and only detect fixed hierarchy structures that are not consistent with real-world communities; methods with hand-crafted thresholds bring difficulties in real applications, while choosing the community corresponding to the maximal belonging coefficient for each node results in a single boundary and neglects the multi-resolution of communities. In order to solve the limitations above, we propose a novel structure to dig finer information by partitioning the members into several levels according to their belonging coefficients. We call this novel structure Hierarchical Structure of Members (HSM) and discuss its properties in continuity, comparability, consistency and stability which reveal the multi-resolution of community as well as the intra-relations among members. We propose a two-phrase method, Random Walk and Linear Regression (RWLR), to detect HSM. The method measures the belonging coefficients of members by random walk and then divides the members into multiple segments by linear regression. Experiments show that members in the same level hold the same properties and HSM reveals multi-resolution of community. Besides, the comparison in benchmarks shows the efficiency in community detection. Finally, we apply HSM to analyze social networks, including visualization of community structures in large social networks and interactive recommendations in Amazon network.