Research Of Network Embedding Based On Deep Learning

With the rapid development of information technology,network has become an important form of data.Analyzing network data can benefit many realistic tasks such as node clustering,social recommendation and community detection.The sparsity and great scale of network make it not easy to save and handle data.So,network embedding is proposed to solve the difficulty of data storage and processing.Network embedding aims to learn a low-dimensional and dense representations for every node in the network.With the continuous development of deep learning,the performance of network embedding algorithms based on deep learning is far superior to traditional network embedding algorithms.However,the existing network embedding algorithms do not make good use of the label and attribute information of the nodes to learn the underlying characteristics of the network.Therefore,according to the different characteristics of labeled network,attribute network and attribute label network,three network embedding algorithms based on deep learning are proposed in this paper.The main research contents of this article are as follows:(1)A network embedding algorithm considering global structure and label is proposed.Distance metric matrix which preserves the global structural information of network is constructed based on the Dijkstra algorithm.Denoising autoencoder is used to capture the high non-linear characteristic of the network topology.The output of the encoder is the learned representations and can be used as the input of the side output network layer to predict the category of the node.Label information is preserved by minimizing cross-entropy loss and center-loss.Experimental results show that in node classification and visualization tasks,the proposed algorithm has a significant improvement over the baseline algorithms.(2)A network embedding algorithm considering community information and attributes is proposed.The proposed algorithm exploits a deep model based on encoder-decoder architecture and early fusion strategy to capture the complex interaction between topological information and node attributes,uses Louvain algorithm to divide the network into multiple communities,construct community triplets,and uses the output of encoder,which is the learned representations,and community triplets to calculate the triplet loss.Experimental results show that in node classification,node clustering and visualization tasks,the proposed algorithm is better than the baseline algorithms.(3)A high-order network embedding algorithm considering label and attributes is proposed.The proposed algorithm exploits node attributes to construct an attribute graph and an attribute masking matrix,and constructs multiple masking matrices based on the shortest distance between nodes.The abovementioned masking matrices are used in the attention mechanism to extract the first-order neighborhood information of the attribute graph and multi-order neighborhood information of source structural graph to form multiple node representations.Attention mechanism is used to adaptively determine the contribution of different node representations.Experimental results demonstrate that our method has a higher accuracy rate on classification tasks.