Node Classification Method In Social Network Based On Graph Encoder Network

The rapid rise of social platforms has brought a lot of social network data,which contains rich information.How to mine the effective information has attracted the attention of scholars.Social network node classification is one of the research hotspots.However,the existing node classification methods in social network still face the following challenges:(1)How to flexibly integrate node attributes and network structure of the social network,and mine the implicit interaction between nodes.Most of the existing algorithms are based on traditional classifiers or label propagation methods based on random walks,which cannot take into account the entity attributes and interactions of social networks,and rarely consider the interaction differences between different entities;(2)How to realize the node classification of temporal social network.Social networks are often dynamic,how to combine the temporal information and spatial information of the social network to realize the classification of the temporal network is still a problem.In view of the above problems,we proposed a node classification algorithm in social network based on graph encoder network.The main work of the paper includes:1)In order to flexibly integrate the nodes' attributes and structure of the social network,mine the interaction information between nodes,a node classification method in static socail network based on graph encoder network is proposed.The graph encoder network regards entities as nodes,and now considers a entity as a target node.First,the target node propagates its information to its neighborhood;Then,in order to consider the differences between entity relationships,we further mine the implicit interaction information between the target node and the neighbor nodes,and aggregate the interaction information;Finally,updates the node's fearure according to the its aggregated interaction information and its attributes,and predict the node's label according to this feature.We verified the effectiveness of the model by experimenting on the Weibo dataset and the DBLP paper citation dataset.2)In order to realize the node classification in temporal social networks,we propose two methods based on graph encoder network for temporal information and spatial information of social networks,which are the node classification algorithm in temporal social network based on spatio-temporal graph encoder network and node classification algorithm in temporal social network based on gated recurrent unit and graph encoder network.The first method assumes that there is an interactive relationship between the current time slice and the previous time slice.However,the information contained in these relationships is often complex and high-dimensional.We use a multi-layer perceptron tolearn the representation of the interaction relationship between time slices,and then combine it with the information of the current time slice to learn the labels of nodes.However,in fact,the behavior pattern of social network entities may have the characteristics of periodicity and large time span,but the first method can not effectively capture the long-term dependent structural information and node information on social network.To solve this problem,we propose the second method,which uses the gated recurrent unit to model the temporal structure information and entity attributes of the social network,and classifies the nodes of the current time slice by combining the information of past time slices.