The Robust Community Detection Approach In Complex Networks With Background Information

This paper aims to explore how to effectively utilize the background information in complex networks for robust community detection.The background information here mainly includes the node attribute information and semi-supervised constraint information in the network.For the above two types of information,this paper designs two different community detection methods respectively.In the first community detection method,we focus on the relationship between network communities and semantic topics in order to steadily discover communities with strong interpretability.This paper introduces the transition probability matrix with a suitable prior into a well-designed nonnegative matrix factorization framework.This new transition probability matrix can well depict the complex relationship between communities and topics.To illustrate the effectiveness of the first method(i.e.Robust and Strong Explanatory Community Detection,RSECD),we conduct rich experiments on both synthetic and real networks.The results show that our new method RSECD is superior to the baselines and some different types of community detection algorithms in terms of accuracy.We also conduct a case study analysis on a musical social network to validate RSECD's strong interpretability to detected communities.The second method mainly uses the semi-supervised information in the network to detect communities.Since Graph Convolutional Network(GCN)and Markov Random Field(MRF)have complementary features,this paper innovatively cast the MRF model to a new convolutional layer and incorporate it as the third(and the last)layer of the GCN model,in order to achieve an end-to-end deep learning method which is also community oriented.To verify the advantage of this second method(i.e.Markov Random Field as Graph Convolutional Network,MRFasGCN),we first quantitatively compare this method with the baseline method GCN as well as the corresponding twostage method on nine real networks of different scales,and then qualitatively analyze some nodes on the Cora dataset which are wrongly divided by GCN and corrected by MRFasGCN.Finally,we compare MRFasGCN with seven representative community detection methods.The results show the superior performance of our new approach.