Network Embedding In Complex Networks

The network is very common in reality world,since it's ability to represent the complex relationships.However,due to the increasing complexity of the network and the large scale of the network,it becomes more and more difficult for the complex network analysis.In order to cope with these problems,researchers have proposed network embedding methods.i.e.,learning low-dimensional vector space from the complex network structure and various attribute information.Based on the embedding vector,we can easily adjust the existing machine learning models to solve tasks in complex network analysis.At the same time,the quality of network representation model will directly affect the final tasks.In complex network analysis tasks,most traditional network representation works capture the structure information by measuring the node similarity.Then most tasks,like node classification/clustering,link prediction,community partitioning can be fulfilled with the node similarity.In the process of measuring node similarity,most methods extract local structure information by the common neighbours,while the high-order information and the complex local structure affecting the node similarity are ignored.Therefore,we propose a similarity measurement method based on graph kernel.This method mainly considers the influence of the relationship between nodes in sub-graphs.And then the similarity measurement method is applied to the sign prediction.The results show that our method has a significant improvement in sign prediction compared to other traditional methods.In the existing network embedding method,network structure information is the basis of all embedding models;however,due to the sparsity of the network itself,the performance of some methods is not good.Therefore,in order to solve the problem caused by the sparseness of network,we propose a novel edge-dual graph based method,which converts the original graph into the edge-dual graph.In the transformed network,we Jaccard coefficients to measure the node similarity,and then apply kernel SVM to predict the node signs.This method not only improves the sparseness of the network,but also transforms the link-based tasks in the original network into node-based problems in the edge-dual graph.Structural information is the basis of network,and extracting effective information from the complex network is important for the network representation model.In a complex structure,a single structural information acquisition method can only partially reflect the structural information in the network.Therefore,the network embedding model based on single structural information can not obtain the complete network information.Since measuring node similarity can quantify the structural relationship directly,we propose a multi-view network representation model based on integrated similarity to fully exploit the information in the network.According to different node similarity combinations,some potential network representation spaces,i.e.views,are generated for the network.Then,a canonical correlation analysis based method is then used to learn a common representation these generated views,and a characteristic representation of the different views is generated by the neural network.Finally,the common representation and characteristic representation are fused as the complete representation space of the network.As the network becomes more and more complex,both structural information and attributes information are very important for the network.Therefore,the structural information based embedding models result in poor performance on the final task due to insufficient use of information.Although some work attempts to learn network representation space by heterogeneous information,the utilization of the information is still insufficient.Therefore,we propose a novel heterogeneous information network embedding model(HINE).The HINE model makes full use of user attributes,links,community structure information,and label information.By fusing a variety of heterogeneous information,results show that our method is better than the existing work performance in the user classification,especially in the low-dimensional space,our method has a superior performance than others.