Research On Link Prediction Method Of Networks Based On Representation Learning

Human activities constitute social groups with different function,these social groups are made up of different people and their intricate relationships.With the rapid development of the Internet and modern information technology,the traditional social groups are gradually transfering and developing online,and online social network has become the main platform for carrying human activities.As an important dependence of people's daily life,social network has numerous nodes and complex information,which constantly evolves over time,these making it increasingly difficult to study.Link prediction,as the crucial research content in social network analysis,is helpful to explore the evolution mechanism and development mode of the network,evaluate social phenomena,improve the missing information in the network,and dig the potential links among users.This thesis takes the current mainstream social platforms as example and combines the network representation learning technology to study the future links and unknown links.The main research work of this thesis is as follows:1.In terms of the prediction of future links,from the perspective of the heterogeneity of features in social networks,a link prediction model based on feature representation and fusion is constructed.Firstly,based on the sparseness and high-dimensionality of network structures,network embedding is applied to represent the network structures as low-dimensional vector,which can identify the spatial relationship and discover the relevance between users.Secondly,owing to the diversity and complexity of text semantics,word embedding model is introduced to vectorize user text,and time decay function is introduced to quantify the influence of user text on link establishment.Meanwhile,the top-k relevant users of each user are selected to simplify the computational complexity.Finally,a link prediction method based on Attention and feature fusion is proposed,which fuses and mines structural features and user text features,and finally realizes the purpose of synthetically predicting links from multiple feature Spaces.The experiment shows that the model can effectively improve the performance of link prediction.2.In terms of the prediction of unknown links,from the perspective of the diversity of features in social networks,fully considering the correlation between different factors,this thesis proposes and designs a link prediction model based on improved network representation learning.This model explores the potential association among users,and finds out the key factor affecting the establishment of links.First of all,based on user relationship and user behavior,the matrix of common friend ratio,the matrix of common behavior ratio and the matrix of interest similarity are constructed respectively.Then using the advantage of deepwalk algorithm in network representation,the three matrices are fused and used as the transition probability of random walk in deepwalk to obtain the vector representation of users.Finally,the similarity between vectors is calculated and the threshold value is estimated to predict the unknown links between users.The experiment indicates that the proposed model can effectively predict the links between users and find the key factor of link establishment.