Research On Multi-granularity User Alignment Across Social Networks

In recent years,with the rapid development of the Internet and social networks,in order to enjoy all the services provided by the social network platform,users will register accounts on multiple social network platforms.User alignment is the discovery of the same person's account on multiple social networks,which is important in many applications.In terms of user alignment,there are currently two main challenges.First,the cost of collecting manually aligned users as training data is very high,but traditional supervised methods often require a large amount of labeled data to obtain better results.Second,Different social networks have different functions and services,resulting in different user structures and attribute types,which further increases the difficulty of user alignment.In order to solve these two problems,this thesis attempts to combine the idea of multi-granularity calculation and propose an unsupervised user alignment method based on multi-granularity calculation.The main research results are as follows:1.An unsupervised user alignment method SPUAL based on user attributes and structure is proposed.A novel soft alignment consistency principle based on user attributes and structure is designed.Different soft alignment principles are assigned different weights.The method calculates the degree of alignment consistency between user pairs.It is assumed that aligned user pairs should meet certain soft alignment consistency principles.Based on this assumption,the objective function is designed and converted into a matrix to obtain the alignment matrix.Finally,a greedy matching method Find out all aligned users.Experiments performed on several public data sets show that the alignment accuracy of this model is significantly improved over the current most advanced unsupervised methods.2.An unsupervised user alignment method MGUAL based on multi-granularity is proposed.First,the multi-resolution matrix is used to roughen the social network.The coarsened adjacency matrix of the network can capture the internal structure of the social network.After the coarsening,the classic unsupervised user alignment method is used to obtain the current granularity alignment matrix,and the interpolation matrix is used to analyze the relationship between the alignment matrices of different granularities to obtain the finest granularity alignment matrix.With this method,the time complexity of the model can be significantly reduced without losing alignment accuracy.Experimental results on multiple data sets prove that the MGUAL model has high time performance and can accurately identify aligned users.