Research On The Inference Methods Of Social Network User Attributes

Social network user attribute inference refers to inferring undisclosed attribute information of users(such as personality tendencies,family locations,interests,etc.)from social network platform data.These attribute information can provide reliable data support for personalized services,Internet supervision,big data analysis and other fields.Social network user attribute inference has always been widely concerned by industry and academia.Existing methods generally obtain the characteristic description of user information through statistical analysis,embedded representation and other methods,and then use supervised learning,regression analysis and other means to infer the user's attribute category or attribute value.However,the existing methods still have the following two problems in practical application: 1)In terms of information utilization,the multi-faceted information related to user attribute inference and the correlation between these information are not fully considered.2)In terms of feature representation,classical feature representation methods(such as statistical analysis,embedding models)cannot accurately represent user features.Both of these problems lead to the low final accuracy of user attribute inference.To this end,this thesis studies the method of social network user attribute inference,and the main contributions are summarized in the following two aspects:(1)A user tendency inference method based on knowledge graph meta-path is proposed.This method utilizes the characteristics of knowledge graph,which has the characteristics of efficient organization and good interpretability of heterogeneous information,establishes the connection between users and various information in social networks,realizes the utilization of various information and considers the correlation between information.In addition,based on the knowledge graph,this method designs a meta-path walk algorithm to extract multi-dimensional walk features,which can be used to represent users accurately.Experiments show that the method proposed in this thesis has achieved good results in information utilization,feature representation,tendency recognition and so on.(2)A user location inference method based on multimodal information fusion is proposed.This method comprehensively selects three aspects of modal information related to the user's location,and calculates the correlation between these three aspects of modal information based on tensor fusion theory,which solves the problems of insufficient information utilization and insufficient consideration of the correlation between information.In addition,this method builds a multi-modal information fusion model,which processes three modal information through the information compression module,the tensor fusion module and the location inference module,and can obtain more accurate user characteristics.Through the actual data test,the method can fully utilize the multi-faceted information related to the location and consider its correlation,and achieve good results in user feature representation and location attribute inference.