Abstract
With the technological development, the volume of collected data has been increased very rapidly. The corresponding graphs become much larger and more complex. When we process the large-scale networks, the most existing methods [1] do not work. The reason is that most of the algorithms need to trace the whole network for each step, but the network is too huge to handle. Both the time complexity and the computational complexity have grown up rapidly. For this reason, we are interested in developing a community detection algorithm for solving the large-scale networks without tracing the whole network for each step. In our thesis, we define the core of a set that can be used to represent the set. In large-scale networks we can ignore the set S and focus on the core of the set S during the process of community detection. We propose an algorithm, called the core-based local community detection algorithm, to verify that the core of a set can represent the set, and test the algorithm by using the LFR benchmark graphs and the DBLP co- authorship network. The core-based local community detection algorithm performs well in the LFR benchmark graphs. For communities with strong community strength in the LFR benchmark graphs, this approach could reach almost 100% precision and 100% recall. However, the core-based local community detection algorithm does not performs well when the communities of the networks have overlapping communities. Finally we develop a core-based community detection algorithm for large-scale net- works, which needs not trace the whole network for each step, and also applies the algorithm to the LFR benchmark graphs and the DBLP co-authorship network. We also compare with three different methods, the label propagation, the fast unfolding and the greedy optimization of modularity, respectively.