Research On Accounts Matching Method For Cross-Online Social Network Platform

At present,the user scale and data scale of OSN(Online Social Networks)are expanding rapidly,which gives rise to a variety of cross-social network applications.Among them,accounts matching across social network platforms has gradually developed into a hot research field.The existing researches on accounts matching across social network platforms use a variety of user data to build relevant models to depict user portraits,so as to establish users matching schemes.However,there are still three problems in the existing methods: 1)the traditional methods based on multi-source user data integration are limited to the multimodal,fragmented,asymmetric and other characteristics of user data,which face the challenge of effectiveness and practicality;2)the existing methods deal with different user features alone,which can not effectively integrate user multimodal features,and the matching accuracy needs to be improved.3)the traditional cross-social network accounts matching methods focus on the extraction of user features while ignoring the efficiency of users matching.Aiming at the above three problems,this thesis studies the user data and feature extraction and matching scale,and puts forward the following three points to solve the current problems of cross-social network accounts matching.In order to solve the problem of difficult coupling of spatio-temporal information and difficulty of feature extraction caused by the difference between user check-in and traditional trajectory in social networks,a similarity measurement method based on user check-in is proposed.First,grid mapping of the check-in data is carried on,and association filling algorithm to generate the user check-in tensor is proposed to retain the check-in information to the maximum extent and strengthen the association feature extraction at the same time.Then a feature extraction model is built,residual module and self-attention module is used to learn the correlation of check-in data,the ability of spatio-temporal feature extraction is enhanced,and different features through deep separable convolution is fused.Finally,the multi-layer perceptron is used as the classifier to learn the potential relationship between feature data and user matching,and the user’s score for similarity is achieved.Experimental resuilts show that the method used in this thesis can achieve more accurate similarity measurement.In order to solve the problem that social network node embedding can not effectively integrate user multi-dimensional information,a node embedding method which combines user characteristics,network topology and friend relevance is proposed.Based on the graph attention network,the node is projected into the hidden space,the node correlation coefficient is calculated by the attention mechanism,and the new node features are generated and embedded into the node vector through the connected multi-layer perceptron.Finally,vector coding which combines the three kinds of information of node characteristics,network topology and friend relevance is obtained,so that the classification of user interests topics can be realized more accurately.Experimental results show that the node embedding method proposed in this thesis achieves good results in each dataset.In order to solve the problem of low efficiency of cross-social network users matching,this thesis proposes a stable node matching method based on LSH(Locality Sensitive Hashing)multi-bucket strategy and GS(Gale-Shapley)algorithm.First,the user storage structure in the bucket is constructed by node embedding and three-dimensional meshing,and the user search strategy in the bucket is optimized to reduce the matching overhead.Then the set of candidate matching users is obtained by in-bucket search and similarity measurement model and the matching chain is constructed.Finally,the matching chain is clipped by GS algorithm to obtain stable matching results.Compared with existing related research,the users matching method proposed in this thesis can effectively reduce the matching overhead and achieve good results under the relevant indicators.Through experimental tests of three sets of data sets,results show that the proposed method can improve the efficiency of crosssocial network accounts matching on the premise of meeting the users matching index.