Unfolding network communities by combining defensive and offensive label propagation

Abstract: Label propagation has proven to be a fast method for detecting communities in complex networks. Recent work has also improved the accuracy and stability of the basic algorithm, however, a general approach is still an open issue. We propose different label propagation algorithms that convey two unique strategies of community formation, namely, defensive preservation and offensive expansion of communities. Furthermore, the strategies are combined in an advanced label propagation algorithm that retains the advantages of both approaches; and are enhanced with hierarchical community extraction, prominent for the use on larger networks. The proposed algorithms were empirically evaluated on different benchmarks networks with planted partition and on over 30 real-world networks of various types and sizes. The results confirm the adequacy of the propositions and give promising grounds for future analysis of (large) complex networks. Nevertheless, the main contribution of this work is in showing that different types of networks (with different topological properties) favor different strategies of community formation.