The task of community detection in attributed networks is to find a good community partition in terms of both topology structure and node attribute. Despite that a number of algorithms have been suggested for community detection in attributed networks, most of them suffer from considerable performance deterioration when the community structure is not clear or the attributes of nodes in one community are not homogeneous. In this paper, we suggest a dual-population-based multi-objective evolutionary algorithm, called DP-MOEA, for balancing topology structure and node attribute in community detection of attributed networks. In DP-MOEA, one population takes charge of community detection according to the topology structure, whereas the other population is responsible for community detection based on the node attribute information. The two populations evolve independently by different genetic operations and interact with each other at every certain number of generations to utilize the good individuals obtained in the other population. Moreover, a node attribute similarity-based local search strategy and a community merging strategy are designed in the procedure of population interaction to enable the generation of high-quality individuals. Experimental results on synthetic and real-world attributed networks demonstrate the superiority of the proposed DP-MOEA over eight state-of-art evolutionary algorithms for community detection in attributed networks, especially when the community structure is unclear or the node attributes in one community are not homogeneous.

属性网络中社区检测的任务是在拓扑结构和节点属性方面找到一个好的社区划分。尽管已经提出了许多算法用于属性网络中的社区检测，但当社区结构不清晰或一个社区中节点的属性不均匀时，它们中的大多数都会遭受相当大的性能下降。在本文中，我们提出了一种基于双种群的多目标进化算法，称为 DP-MOEA，用于在属性网络的社区检测中平衡拓扑结构和节点属性。在DP-MOEA中，一个种群根据拓扑结构负责社区检测，而另一个种群根据节点属性信息负责社区检测。这两个种群通过不同的遗传操作独立进化，并在每隔一定代数相互作用以利用在另一个种群中获得的优良个体。此外，在种群交互过程中设计了基于节点属性相似性的局部搜索策略和社区合并策略，以生成高质量的个体。合成和真实世界属性网络的实验结果证明了所提出的 DP-MOEA 优于八种最先进的属性网络社区检测进化算法，特别是当社区结构不清楚或一个社区中的节点属性不明确时不均匀。