Research On User Identification Methods Across Social Networks

In the era of mobile Internet,people's social needs on the Internet are increasing day by day,and various social platforms have emerged as the times require,and the number of users and accounts has exploded.The user identification technology across social networks can identify the user's identity by processing and analyzing the user's account information,friend relationship,content and other information,and has important value for friend recommendation,advertising,national defense and security and other fields.Existing user identification methods generally obtain the vectorized representation of the user's identity through statistical information or feature representation,and then use various mathematical models to identify the user's identity.However,there are still two problems in the existing methods:1)In terms of information utilization,the multi-faceted information related to user identification and the correlation between these information are not fully considered.2)In feature extraction,the existing methods fail to extract some targeted features based on the user's identity information.In order to solve the above problems in existing methods,this thesis selects two subdivisions of account matching and attribute identification in user identification for research.The main contributions are as follows:(1)An account matching method based on multimodal feature tensor fusion is used.This method selects the information of four modalities related to user identity,and uses the method of tensor fusion to capture the correlation between the information of each modal,which solves the problem of insufficient information utilization and establishes the correlation between the information.sex.In addition,the method uses personal profiles to construct a heterogeneous graph of user information,and uses an improved random walk algorithm to extract features,which can more accurately characterize user identities.Experiments show that theF₁ value of the account matching method used in this thesis is 1.34%higher than that of the traditional method.(2)A user attribute recognition method based on contextual features is used.In addition to the regular word embedding features,the method formulates a large number of targeted contextual features for user attributes,introduces character-level embeddings for possible misspelled words in tweets,and uses a bidirectional long-term and short-term memory network.Work units are identified using a combined model with a conditional random field.Experiments show that theF₁ value of the user attribute recognition method used in this thesis is 10.30%higher than that of the traditional method.