Research On Network Representation Learning Based On Graph Attention Model

In the era of big data,the connection between things is mostly presented in the form of network.The network widely exists in people’s real life,such as personalized recommendation in recommendation system,community discovery,traffic prediction,road planning and community detection.How to mine the information in these networks is the current research hotspot.In real life,the nodes and links that make up the network usually change over time.With the addition,deletion and change of these network links and nodes,the network also changes over time.However,at present,most of the research on network representation learning focuses on the static network,ignoring the time dynamics of the network to a great extent.This paper proposes dynamic network representation learning methods based on graph attention neural network and temporal attribute information mining.Aiming at the dynamic representation learning of time-series networks,this paper first proposes a graph attention network-generative adversarial networks(GATGAN)model based on time-series information.The model adopts the depth framework base Generative Adversarial Networks.Firstly,graph attention network is used to encode the depth nonlinear characteristics of nodes in the network,and then combined with long short-term memory network and attention mechanism to extract and retain the key time series information in the graph and capture the time dependence of the whole dynamic network.Through the joint learning of confrontation module and graph attention model,the nonlinear spatiotemporal features in dynamic networks can be captured at the same time.Aiming at the dynamic representation learning of attribute networks,this paper proposes a graph attention network-auto-encoder(GAT-VAE)model based on attribute information.Firstly,variational auto-encoder is used to capture the low dimensional potential features,and then a network encoder with multiple graph attention layers is designed to capture the network structure and node attributes.The structure decoder is used to reconstruct the original network structure,and the attribute decoder is used to reconstruct the original node attributes.Finally,long short-term memory network is used to learn the timing information of each node connection state and fuse the time attributes of the network.This method can effectively learn the attributes of network nodes.Aiming at the problem of wind speed prediction,this paper proposes a wind speed prediction method based on a graph attention network-auto-encoder(GAT-VAE)model.Firstly,variational auto-encoder is used to reduce the dimension of feature space,and multiple wind farms are constructed as spatial network structure.Then the spatial correlation of time series data is learned through graph attention network,and the temporal correlation of meteorological data is extracted by long short-term memory network.The experimental results show that the wind speed prediction model has good prediction effect.