Research On Social Network Representation Learning

With the rapid development of information technology,social networks have gradually become an essential part of daily human life.Recently,Internet data has grown exponentially in recent years,and the social networks have accumulated a large amount of meaningful data.However,the relationships between nodes are becoming more and more complicated,which brings great challenges to the analysis and data mining of the social networks.How to perform network analysis tasks in a large amount of data to mine potential values is the critical problem that needed to be solved.Social network representation learning,also known as social network embedding,namely,embedding the nodes of a social network into a low-dimensional space,which is the foundation in the analysis task of social networks.Recently,many methods based on graph convolutional networks have been proposed to solve the problem of network embedding,These algorithms assigns the same weight to all neighbor nodes,and obtains the representation of the central node by aggregating the characteristics of neighbor nodes.However,it ignores the difference in the degree of association of node features,decreases the node features with high correlation and enhances the node features with low correlation,and the features with low correlation will interfere with the generated node characterization,so there is still some space for the improvement about the obtained node representation quality.In addition,there are some noisy nodes in the social network in reality.The noisy nodes have obvious differences from other nodes in the same cluster,which has a negative impact on the quality of their neighbors.In order to solve the above problems,this thesis considers Homogeneous attributed social networks for the research,and the specific research content of the design algorithm model is as follows:This thesis designs the NSGAE model,which is based on a graph neural network,evaluates the importance of neighbor nodes by calculating the degree of association between node features,and uses neighborhood sampling to select a fixed number of highly correlated nodes for aggregation operations to enhance the nodes Indicates quality.The algorithm model is composed of two parts: encoder and decoder.In the encoder part,the intermediate result of network embedding is obtained by stacking multi-layer graph convolution operation.The decoder part trains the entire model by reconstructing the links between nodes.After the model converges,the low-dimensional vectors obtained from the middle layer are represented as nodes,and the network embedding results are obtained.This thesis proposes the NRSNE algorithm based on NSGAE.The algorithm imposes the reconstruction target neighbor constraint on the self-encoder model,making the noise node's feature representation depend on its neighbor nodes,reducing the negative impact of the noise node,and obtaining attribute features.In order to reduce the information loss caused by neglecting some partial links in the neighborhood sampling process,the model is based on the principle of homogeneity and integrates the denoised attribute features with the features in the NSGAE model to generate new node representations.In the experimental part,this thesis compares the result of multiple groups of existing algorithms on node classification,link prediction and clustering tasks,and then,evaluates and analyzes the quality of node representation.The experimental results show that,considering various information and modeling of the same network,the NSGAE algorithm can effectively enhance the quality of node representation and improve the effect of node representation in subsequent clustering and classification tasks.The NRSNE algorithm can effectively avoid the influence of noise nodes,and has a prominent advantage in classification,clustering and link prediction tasks.At the same time,it has a stable advantage in relatively sparse social networks.