Research On Link Prediction In Social Networks Using Attribute Inference

Social media has become an increasingly important resource for connecting with people,processing information,and expanding social influence.When using social media,social media users want to be recommended content or friends of interest.However,they do not want to provide personal information and disclose their existing friendships due to privacy concerns.Therefore,attribute inference and link prediction have become important solutions to this problem.The essence of attribute inference is to identify the missing attributes of nodes in a social network.The essence of link prediction is to predict whether a link is likely to be established between two nodes in the network.Currently,most studies on attribute inference have been conducted based on the structure of social networks or users’ own behaviors and rarely take into account the similarities that exist among users themselves.Apart from that,most of the current research separates attribute inference and link prediction and studies them as two different works.However,there is a strong correlation between the two works both in terms of their nature and application scenarios.For the above problems,the following research is conducted in attribute inference and link prediction in this thesis:(1)Attribute Inference Based on User Similarity and Random Walk(USRW)is proposed.First,the algorithm calculates the user’s score for possible missing attributes through a collaborative user-based filtering method.Then,a restarted random walk is performed in a weighted heterogeneous information network built based on the social network.The user’s score on the attribute is calculated using the network structure and the user’s existing social relationships.Finally,the attribute scoring results obtained by the two methods are combined for attribute inference,and experiments are conducted on the publicly available dataset Deezer Social Networks.It is found that the proposed algorithm in this thesis incorporates user similarity and network structure factors,and the inference accuracy is greatly improved compared with the traditional algorithm.(2)A Social Networks Link Prediction Based on Graph Convolutional Neural Networks(GCN-SNLK)is proposed.First,attribute inference based on user similarity and random wandering is performed.The attribute inference results are incorporated into a user-attribute network.Later,using a multilayer graph convolutional network,the multiorder connectivity between user attributes in the network is modeled and used to learn an embedded representation of the user.The learned embedded representations are entered into a multilayer perceptron to predict the probability of the existence of links between users in a social network.Finally,the proposed algorithm is experimented on the publicly available dataset Deezer Social Networks.It is found that the proposed GCN-SNLK model improves in terms of normalized cumulative discount gain and hit rate compared with the traditional link prediction algorithm.(3)The proposed attribute inference and link prediction algorithm algorithms are applied to design and implement a music recommendation system based on attribute inference and link prediction.The system adopts Attribute Inference Based on User Similarity and Random Walk to infer the attributes of users and recommend music and artists that may be of interest to them.The graph convolutional neural network-based social network link prediction algorithm to recommend users or potential friends that may be of interest to them.