Influence-Based Community Detection and Ranking

A Leader-follower algorithm is proposed to detect the communities of varying sizes. Leaders are determined using distance centrality and followers are associated with their leaders using cosine similarity method. The entities are ranked by a combination of their local (PageRank) and global (Betweenness Centrality) importance in the network. The communities formed are admissible if the calculated modularity is 0.3 or higher. The Leader-follower approach outperforms similarity-based clustering methods in terms of modularity values. From the ranking algorithm, the top-ranked entities were interpreted as entities with membership across communities and low-ranked entities were interpreted as outliers (using Gephi as the visualization tool).