Network Embedding Model Based On Edge Sampling

With the development of large-scale on line social networks such as Weibo,WeChat,and Facebook,network representation learning(network embedding)has aroused widespread concern in academe and industry.Network-based studies,such as network node classification,recommendation,anomaly detection,all require a proper representation of network.As the input of machine learning algorithm,the network representation has a very important influence on the learning performance of the algorithm.Traditional network embedding models exploit the spectral properties of various matrix representations of graphs.These models suffer from both computational and statistical performance drawbacks.Recently,a lot of neural network based embedding models are presented in the literature.They have high efficiency and preserve the network structure information well.These models have made remarkable achievements in a number of fields like natural language processing,image recognition,and speech recognition.However,we find that almost all the existing network embedding models only focus on the existence of the edges between nodes,while ignoring the differences between edges.However,the connections between nodes are discriminatively correlated via different relationship types.They imply abundant information.Therefore,we propose NEES,an unsupervised network embedding model.It first optimizes an objective function by edge sampling to obtain the edge vectors,which can reflect the relationship types of the edges.Then we use the edge vectors to improve the existing network embeddings,and learn a low dimension representation for each node in the graph.Empirically,we conduct the experiments on several social networks and one citation network.The results show that our model outperforms the baselines in three applications,including visualization,multi-label classification,and link prediction.In addition,NEES is scalable to large-scale networks in the real world.