Research On Community Detection In Attribute Graphs

With the development of information technology,attributed graph which describes the complex relationships among objects is ubiquitous in the real world,such as social networks,citation networks,protein interaction networks and so on.Community detection for attributed graph is a fundamental and important task,which aims to discover the groups or communities of densely connected nodes.And the problem is how to solve the dual heterogeneity of attribute and topology spaces and improve the performance of community detection algorithms.Graph Convolutional Networks(GCNs)which combine attribute and topology information through convolution operation,are widely used and efficient tools for representation learning of attributed graph.And the research on GCNs is also a hot research field.However,deep GCNs models suffer from over-fitting and over-smoothing and the performance of GCNs models decrease significantly with the number of stacked GCNs layers increases.According to our experiments,the degree of smoothing is different for nodes with distinct degrees in the attributed graph.Inspired by the above experimental results,we propose a degree-specific topology learning method which adjust the topology dynamically according to the degrees of nodes,to alleviate over-smoothing and improve the performance of model on semi-supervised community detection task.The experimental results demonstrate the effectiveness of our method.In addition,due to the high cost of manual labeling,unsupervised community detection for attributed graph is a hot and challenging task.Due to the lack of ground-truth labels,we propose an end-to-end adaptive attributed network embedding model for community detection,and the "soft labels" is generated to guide the model,based on the assumption that node pairs with similar topology and attributes belong to the same community and node pairs with dissimilar topology and attributes belong to different communities.And we design high-order GCNs to introduce high-order information for network embedding.In the meanwhile,we utilize stochastic block model to reconstruct the topology without introducing noise to improve the performance of our model for community detection task.Furthermore,due to the effectiveness of low-pass Laplacian smoothing filter,we propose a symmetric graph auto-encoder model based on low-pass filter,which reconstruct the attributes,introduce self-representation layer and graph mutual information to learn the node embedding preserving the original attribute information with a bit noise.We conduct experiments on three popular datasets and compared with some state-of-art models The experimental results demonstrate the effectiveness of the two models we proposed above.At the same time,in order to combine our proposed methods with practical applications,we design and implement a software system for community detection of attributed graph,which visually shows the results of the methods.