Graph Convolution Based Network Embedding For Link Prediction

Link prediction is an important task and widely used in the field of social network analysis and mining.The graph convolution-based network embedding is an outstanding performance method for link prediction,which learns low-dimensional representations through two steps,node embedding and edge embedding,which can be easily integrated with traditional machine learning algorithms.In this thesis,targeted graph convolutional network embedding link prediction algorithms are proposed separately for networks of different sizes.The main work is as follows:Existing methods often aggregate node embedding through neighborhood iteration when node embedding,which cannot retain global structure information at a lower node aggregation cost;in addition,they fail to analyze the advantages and disadvantages of different embedding methods well when edge embedding,and it is difficult to retain the local structure information obtained by Hadamard product,summation or direct concatenation.To address these problems,this thesis proposes a novel network embedding link prediction method via triadic closure based direct aggregation and weighted concatenation.To improve the efficiency of neighborhood aggregation,the method directly aggregates multi-order neighbors to the central node in node embedding and assigns the neighboring nodes with aggregation weights positively related to their triadic closure neighbors.To better preserve the local structure information,in edge embedding,the method obtains the edge embedding by weighted summation of node embedding based on the analysis of the advantages of the summation concatenation approach,and the concatenation weight of node embedding is positively correlated with their triadic closure neighbors.Extensive experiments demonstrate that the triadic closure based direct aggregation and weighted concatenation enable our proposed approach can better predict the possibility of link existence in small-scale social networks efficiently,outperforming state-of-the-art methods.Existing methods aggregating the information of all neighboring nodes is difficult to achieve when node embedding;meanwhile,when edge embedding,relying on a single channel for concatenation cannot ensure the accuracy of edge embedding information because the information aggregated in the node embedding stage is limited.To address these problems,this thesis a novel network embedding approach for link prediction via decay coefficient based hierarchical aggregation and two-channel concatenation.In order to avoid aggregating all the neighbor node information,the method only needs to aggregate part of the neighbor information in node embedding and assign different sampling ratios to each order of neighbor nodes by introducing the decay coefficients widely used in physics;to compensate for the lack of aggregated information in node embedding,the method utilizes two-channel stitching node embedding features in edge embedding to make full use of different single-channel node embedding.In order to compensate for the lack of aggregated information in the node embedding stage,the method proposes two-channel concatenation node embedding features to make full use of the complementarity between different singlechannel node embedding features and thus improve the accuracy of edge embedding.Extensive experiments demonstrate that the decay coefficient based hierarchical aggregation and two-channel concatenation enable our proposed approach can better predict the possibility of link existence in large-scale social networks efficiently,outperforming state-of-the-art methods.