Research On Interpretable Community Detection Methods In Attributed Networks

Attributed network analysis in real life is a complex problem.In general,the model with simple structure has good interpretability,but poor fitting ability and low accuracy.The model with complex structure has strong fitting ability and high accuracy,but due to the large number of model parameters,complex working mechanism and low transparency,the interpretability is relatively poor.In the community detection task,one is to select the model with simple structure and easy to explain and then train it;the other is to train the optimal model with complex structure and then mine the interpretability of the model.Based on these two different choices,interpretable community detection model can be generally divided into two categories: built-in interpretable community detection and post-interpretable community detection.Through the summary and analysis of the detection work in the existing target communities,the following research results are obtained:Firstly,we propose a Non-exhaustive overlapping community detection algorithm using two-stage Maximum likelihood estimation in attributed networks(Not Mle),in order to learn the connection probability simultaneously using community assignment parameters that describe each community.Attribute information can be captured effectively through attribute embedding,and the combination of attribute and structural information can be encoded using graphs based on the Adjacency Spectral Embedding(ASE)algorithm to make the attributes of nodes with similar structures.Then,a community detection model based on the strength of connections between communities is designed to optimize the impact of structural information on communities,and the node assignment strategy in this model fully considers the properties of real networks(overlap and outliers).Finally,the Maximum Likelihood Estimation(MLE)process is introduced to optimize the objective function effectively to obtain more accurate community detection results.Secondly,a Hierarchical Attention Network(Hi AN)for community detection is proposed,which combines two layers of attention embedding and self-training clusteroriented approach.It effectively performs graph clustering and learning graph embedding in a unified framework using graph structure and node attribute information.The attribute and structural information of nodes in a complex network provide two perspectives for the representation of each node.The attention network model in deep learning is used to encode two kinds of information of nodes.Consistent and complementary information in topology and attribute can be captured for efficient and accurate node representation.Then,we design a unified framework for graph embedding representation and community detection.Learn node classification by integrating node embedding and community detection module in a model aiming at community detection.In the end,abundant experiments are conducted on real-word network datasets,and the results show that the two algorithms proposed in this paper are superior to the existing interpretable community detection methods in attributed networks,which confirms the effectiveness and application value of the proposed algorithms.