Research On User Identity Matching Across Online Social Networks

With the development of Internet technology,online social networks have attracted more and more registered users,and also have shown a diversified trend.The diversity of online social networks greatly enriched the users' online life while also fragmenting the users' identity information.User identity matching across online social networks is a technology for matching multiple social network accounts belonging to the same user in real life,thereby realizing the integration of user information of multiple online social networks,which is of great significance in the areas of commercial recommendation,user information association and network security.Currently,researches on user identity matching across online social networks have made great progress.However,there are still some problems in the researches:(1)It is difficult for the algorithms based on network features to represent the high-order information in the social network.Meanwhile,the lack of utilization of information from other dimensions is another shortcoming.Both of the disadvantages lead to a low precision of the current algorithms.(2)The present algorithms based on the user profile information usually utilize an objective weighting-subjective weight-directed method during the weight calculation period,which made the model's robustness low and the relationship between profile attributes were ignored.(3)The capability for extracting features of the location sequence of the algorithms based on spatio-temporal trajectory was kind of weak,which made it hard to describe the characters of the trajectories accurately.Focused on the problems above,three algorithms of user identity matching across online social networks based on network features,profile features and spatio-temporal trajectories respectively were proposed.The main contents are as follows:(1)An algorithm based on network representation learning(NRL)was proposed.The algorithm was mainly based on network features,and username attribute information was combined to achieve the matching.Because of NRL can represent high-order information in the social network with a low calculation cost which can solve the deficiency of the high calculation complexity and difficulty in calculating high-order information,we introduced this technology to the algorithm.We also merged username attribute information to the algorithm.Finally we got the vectors with information of both network features and usernames and determined matching accounts by the similarity between two accounts.Experimental results in two real social networks showed that the performance of the proposed algorithm can be improved compared with the baseline algorithms.(2)An algorithm which merged the profile information based on fuzzy integral theory was proposed.The algorithm was based on the profile features to achieve the matching.Fuzzy integral theory provides a method to merge multiple attribute information,which can assign weights to attributes reasonably.The theory can solve the problem of weak robustness during the weight calculation to a certain extent.We introduced fuzzy measure and Choquet integrals to the algorithm.Firstly,according to the characteristics of different attributes,we determined different similarity calculation strategies;secondly,we utilized particle swarm optimization method to calculate the fuzzy density of each attribute as the weight;finally,the Choquet integrals between two accounts were calculated as the similarities.Experimental results in three real social networks showed that,compared with traditional algorithms,the F1 score of our algorithms has improved the performance by 3.89% to 25.6%,which verified the effectiveness of the algorithm.(3)An algorithm which based on spatio-temporal trajectory sequence representation was proposed.The algorithm was based on the users' spatio-temporal trajectories to achieve the matching.We introduced recurrent neural network structure to solve the problem of the weak capability of extracting features of the location sequence of current algorithms..Firstly,we preprocessed the data,the coordinates were transformed to small grids and the trajectories were segmented at a certain time interval;secondly,word2 vec method was utilized to mine the semantic information of each location;then,Bi-GRU model was utilized to extract the trajectories sequence features and the vectors of trajectories was calculated;finally,the similarity between two trajectories was calculated to determine whether the two accounts were matching.Experimental results in two location sharing based social networks showed that the proposed algorithm can extract the trajectory sequence features and improve the performance of user identity matching.