Research On Semi-supervised Network Representation Learning Method Of Preserving Community Structure

Network representation learning method,also named as network embedding,can get low-dimensional vector representations of nodes in large-scale networks.While preserving the network structure and node attributes,the nodes are mapped into low-dimensional and dense real value vectors,which can be used as node features to complete network analysis tasks such as node classification,clustering,link prediction and visualization,etc.In recent years,scholars have put forward many efficient network representation learning models,most of which only consider the micro topological structure of the network and ignore the constraints of community structure on the similarity of nodes in mesoscopic level;although there are some methods saving community structure for network embedding,most of them are unsupervised models and fail to effectively use node labels or pairwise constraints which can be obtained in advance to improve the effect of node embedding.Most of the existing semi-supervised network representation learning methods use node labels to assist the learning process of node representation vectors,but the use of node labels must know the specific number of communities in a given network,which is more suitable for the task of node classification,dividing different types of nodes more clearly;while the use of pairwise constraints information only needs to know the association between communities which the nodes belong to,so pairwise constraints are more suitable for network clustering.This work mainly focuses on the semi-supervised network representation learning method which preserves the community structure of network.By effectively using the pairwise constraints information to guide network representation learning process,more discriminative node representation vectors are obtained,they can make the nodes in the same community gather more closely.The experimental results of node clustering,link prediction and network visualization all prove the effectiveness of our method..The innovative achievements of this paper are:(1)A semi-supervised network embedding method based on nonnegative matrix decomposition is proposed.The basic idea of this method is: from the point of view of modifying network topology,the adjacency matrix and node similarity matrix of network are modified by using pairwise constraints information;the modified adjacency matrix is used to maximize the modularity to model the community structure of network,and the modified node similarity matrix is used to preserve the similarity of network topology.The incorporation of pairwise constraints can affect the node representation vectors generated inthe learning process of network representation,which makes the node representations more discriminative in node clustering,link prediction and network visualization tasks.(2)A network representation learning method based on semi-supervised random walk is proposed.The basic idea of this method is: in view of the lack of prior knowledge guidance in the sampling process of traditional random walk sequences,our work uses the guidance of pairwise constraints information to affect the neighborhood structure of the current node during random walk,and then adjust walking probability of the next node.The integration of prior knowledge makes the learned node representation more discriminative.The experimental results on node clustering and network visualization tasks show that using pairwise constraints information to guide random walk is helpful for conventional Skip-gram model to learn more discriminative embedding vectors.