چكيده به لاتين
In social networks there’s a set of members which has more strong relationships which is called a community or a cluster which represents valuable information about the type of relations between members, the type of transferring information and the kind of distribution of these members in social networks. The usage of community detection is in social network analysis, electronic commerce, decision support systems, link prediction, natural language processing and medical sciences. Several community detection methods have been proposed for clustering networks. However, none of them is ideal method for network clustering. Therefore, it seems a good idea to apply a fusion method to get the benefit from strengths of methods and cover their weaknesses. Consensus clustering approach is a fusion method in which a set of community detection algorithms are employed as a consensus and results in a better clustering that leads to increase the stability of communities. The innovation proposed in this thesis consists of two parts. In the first part a new method called "fast projection" is presented for converting bipartite networks to unipartite ones that results in detection of stable and important links and removing the less valuable data. In the second part, a new "consensus approach" (consists of Mitra and Azar algorithms) is proposed to detect communities by employing "fast projection". In this approach a bipartite consensus graph is built then fast projection converts the bipartite network to a unipartite network (in order to combine and compress information and emphasize on important links). After that, one of the base clustering methods detects the communities of the unipartite network. To evaluate fast projection and the consensus approach some popular criteria are used (such as NMI and Modularity) and we examine the quality and performance of the proposed method and base methods. Moreover, we compare our method with another consensus method which is recently proposed. The results of our comprehensive evaluation shows our approach is able to detect a more clear community structure for networks in comparison with base methods and the other consensus method and, it finds network communities more efficiently in most cases. Therefore, our fast projection and consensus algorithm lead to extract valuable relational data, detect qualified communities and effective nodes in graphs.
