基于零模型的社区检测基准网络构造及应用

社区检测对于探索挖掘复杂网络的结构特性具有重要意义，社区检测算法性能对于检测结果具有重要影响。目前用于衡量社区检测算法性能的基准测试网络较为单一，主要包括人工合成网络和真实世界网络。由于真实世界网络中通常缺乏已知社区结构信息，人工合成网络成为衡量算法性能的主要途径，但普遍存在网络微观特性不可调且与真实世界网络差异较大、对检测算法区分度不高、无法更改局部网络结构等问题。为提升人工合成网络性能，该文提出基于零模型的基准测试网络构造方法，首先设计了能够保持中尺度特性的零模型，提升网络微观特性调整灵活度，使其更逼近真实世界网络结构特性；其次设计了能够调整社区结构强弱的零模型，提升网络社区检测的评价准确性；最后设计了能够调整局部拓扑结构的零模型，有效衡量局部社区结构特性变化对于整体网络结构及检测算法性能的重要性。实验结果表明，基于零模型的构造方法能够有效提升基准测试网络的多样性和灵活性，更加逼近真实世界网络特性，因此更能满足对于社区检测算法性能的评价需求，对于提升复杂网络社区检测性能具有重要意义。

Construction and Applications of Benchmark Networks for Community Detection Based on Null Models

Community detection is of great significance for exploring the structural characteristics of complex networks while the performance of community detection algorithm makes important influence on the detection results. At present, the benchmark networks that are used to measure the performance of community detection algorithm mainly include artificial synthetic network and real-world network. Synthetic network has become the main method to measure the performance of the algorithm since the real-world network usually lacks information of known community structure. However, it is found that the microscopic characteristics of the network is unadjusted, which is different from the real-world network, the discrimination of the detection algorithm is not high, and it is inability to change the local network structure. In order to improve the performance of artificial synthetic network, a benchmark network construction algorithm on null-model is proposed. Firstly, an algorithm of null model that can maintain the mesoscale characteristics is built to improve the flexibility of network micro-feature adjustment and make it closer to the real-world network structural characteristics. Secondly, the null model of adjusting strengthen and weakness for community structure is designed for improving the evaluation accuracy of network community testing. Finally, a method based on null model is constructed so as to make some adjustments of the local topological structure for measuring the importance of the change with local community structure characteristics to the whole network structure and the performance on detection algorithm. Experimental results show that the algorithm in view of null model can effectively improve the diversity and flexibility of the benchmark network, thus making the network be more similar with the features of real-world network and meeting the demand for performance improvement of community detection algorithm.