Local Community Detection in Complex Networks

Community structure is an important aspect of network analysis, with a variety of reallife applications. Local community detection algorithms, which are relatively new in literature, provide the opportunity to analyze community structure in large networks without needing global information. We focus our work on a state-of-the-art algorithm developed by Yang and Leskovec and evaluate it on three different networks: Amazon, DBLP and Soundcloud. We highlight various similarities and differences between the geometry and the sizes of real and annotated communities. The algorithm shows robustness to the seed node, which is also demonstrated by its rather high level of stability. By using two different methods of seed selection from the literature, we demonstrate further improvement on the quality of the communities returned by the algorithm. Finally, we try to detect reallife communities and show that the local algorithm is comparable to global algorithms in terms of accuracy.