Convergence improvement of differential evolution for community detection in complex networks

Abstract

To improve the quality of optimal partitions obtained by modularity optimization in community detection of complex networks, we summarize two key factors determining the convergence performance of the EA-based (Evolutionary Algorithm) optimization algorithms and present a new Classification-based Differential Evolution algorithm for Modularity Optimization (CDEMO). On the one hand, CDEMO redesigns the main evolutionary operators of the standard Differential Evolution (DE), including the mutation, parameter adjustment and selection strategy, to improve the global convergence ability of the optimization strategy, which is often been overlooked in available EA-based modularity optimization algorithms. On the other hand, CDEMO improves the community modification method to better utilize the known topology information of networks, which reduces the search space of DE and ensures adequate space for the global optimum at the same time. The performance of CDEMO is evaluated on both artificial computer-generated and real-world social networks, and experimental results prove the validity of the improvement measures and the superiority of CDEMO over several existing state-of-the-art modularity optimization algorithms.