Research And Implementation Of Network Embedding Method Based On Deep Learning Model

Nowadays,network as an important carrier of data,is becoming more and more complex with the exponential growth of Internet data.In a network,the relationship between nodes also has huge evaluation of data mining,besides the information contained in the node itself.The goal of network embedding is to solve the limitations of traditional network analysis technology,which is applied to complex machine learning and data mining tasks in large-scale network.The network embedding method based on matrix factorization can reconstruct the network well,but it is easy to lead to over fitting,and the overall performance for other machine learning tasks is not satisfactory.The network embedding algorithms based on natural language model are able to use network structure in different degrees to learn effectively embeddings of nodes,but they all belong to the shallow layer which means that it is difficult for them to learn the deeper and more complex features of the network structure.Deep learning has developed rapidly in recent years and has made important progress in many fields.The essence of Deep learning is to abstract the data features deeply and learn the mapping function from high-dimensional vector space to low-dimensional space.Network embedding can also be regarded as the process of converting the node representation from the high-dimensional space of the original network into a low dimensional vector space whose essential problem is to learn the mapping function between these two vector spaces.Consequently,there are some network embedding methods based on deep learning.However,most of the existing methods based on depth learning only use network structure for network representation,without known labeling information and attributes of nodes,so that they can not better reflect the authenticity of nodes.Therefore,this paper focuses on deep learning based network embedding.The main work is as follows:(1)In view of the lack of robustness to the noise of the stacked auto-encoder(SAE)and the shortcomings of the existing model's that these methods do not exploit known labeling information,a semi-supervised network embedding methods based on contractive auto-encoder is proposed,namely LSDNE.On the basis of the SDNE model,LSDNE replaces stacked auto-encoder SAE with the contractive auto-encoder CAE,and the SVM classifier is used as the supervised part of the model,so that the information of the known label can be incorporated into the process of network embedding.Then,experiments on Citeseer datasets and Cora datasets show that contractive auto-encoder CAE makes the generalization of LSDNE better than SDNE,while the LSDNE model has better accuracy of label prediction than the state-of-the-art techniques.(2)In the real-world networks,nodes contain a variety of attributes in addition to the network structure and labels,which can be used as side information in the process of network embedding.When LSDNE is applied to the attribute network,a large number of super parameters will be generated,which greatly increases the complexity of the model and thus reduce the model training efficiency.In order to make better use of the attributes of network vertexes and reduce the complexity of the model,inspired by the LLE(Locally Linear Embedding)algorithm,semi-supervised network representation learning model SILDNE based on the neighbor structure,namely SLLDNE(Structural Labeled Locally Deep Nonlinear Embedding).The experiments results show that the performance of SLLDNE is equivalent to LSDNE,with fewer parameters.On the basis of SLLDNE,this paper proposes SILDNE further which exploits the known labeling information.The SVM classifier is adopt to classify the known labels,which integrates network structure,labels and node attributes into deep neural network.The experiments results show that SILDNE substantially outperforms the state-of-the-art methods on the multi-label classification task.