Multi-similarity Co-training Spectral Clustering For Community Detection On Dynamic Networks

As one of the fundamental characteristics of complex networks,community structure is very important in the study of complex systems.As the real network is dynamic,dynamic community detection methods have gained wide attention in recent years.In this paper,a multi-similarity co-training spectral clustering method is proposed for community detection on dynamic network.This method firstly consider multiple similarity matrices by co-training and then consider historical information with the idea of evolutionary clustering.Finally,the proposed dynamic community detection method is applied to the validation of artificial data sets and real networks to verify the validity and accuracy.The main work of this paper is as follows:(1)A multi-similarity spectral method is proposed as an improvement to the former evolutionary clustering method.The similarity measurement between data points has not formed a consistent standard.These metric can be viewed as the mapping function of data points from low dimension to high dimension,different measurement methods can lead to different clustering results.A co-training algorithm is introduced by bootstrapping the clustering of different similarity measures,which reduces the sensitivity of the spectral clustering algorithm to the construction of similarity matrix.(2)As the connection between the network nodes and the nodes is not always static in the real complex network,and the social network is dynamic.The two step strategy and evolutionary clustering are used in the existing dynamic community detection algorithms.With the idea of evolutionary clustering,a dynamic co-training algorithm is proposed for dynamic networks.(3)In order to apply better to the real network,a method is proposed to adaptively select the smoothing parameters in evolutionary clustering.And the proposed method is extended to the case that the number of community and node is changed with time.(4)Compared with a number of baseline models,the experimental results show that the proposed method has better performance on some widely used synthetic and real-world datasets with ground-truth community structure that change over time.