Community Detection in Complex Networks

Abstract

Finding communities of connected individuals in social networks is essential for understanding our society and interactions within the network. Recently attention has turned to analyse these communities in complex network systems. In this thesis, we study three challenges. Firstly, analysing and evaluating the robustness of new and existing score functions as these functions are used to assess the community structure for a given network. Secondly, unfolding community structures in static social networks. Finally, detecting the dynamics of communities that change over time. The score functions are evaluated on different community structures. The behaviour of these functions is studied by migrating nodes randomly from their community to a random community in a given true partition until all nodes will be migrated far from their communities. Then Multi-Objective Evolutionary Algorithm Based Community Detection in Social Networks (MOEA-CD) is used to capture the intuition of community identi cation with dense connections within the community and sparse with others. This algorithm redirects the design of objective functions according to the nodes' relations within community and with other communities. This new model includes two new contradictory objectives, the rst is to maximise the internal neighbours for each node within a community and the second is to minimise the maximum external links for each node within a community with respect to its internal neighbours. Both of these objectives are optimised simultaneously to nd a set of estimated Pareto-optimal solutions where each solution corresponds to a network partition. Moreover, we propose a new local heuristic search, namely, the Neighbour Node Centrality (NNC) strategy which is combined with the proposed model to improve the performance of MOEA-CD to nd a local optimal solution. We also design an algorithm which produces community structures that evolve over time. Recognising that there may be many possible community structures that ex- plain the observed social network at each time step, in contrast to existing methods, which generally treat this as a coupled optimisation problem, we formulate the prob- lem in a Hidden Markov Model framework, which allows the most likely sequence of communities to be found using the Viterbi algorithm where there are many candi- date community structures which are generated using Multi-Objective Evolutionary Algorithm. To demonstrate that our study is effective, it is evaluated on synthetic and real-life dynamic networks and it is used to discover the changing Twitter communities of MPs preceding the Brexit referendum.