The Design And Implementation Of User Profiling Across Networks Based On Transfer Learning

With the development of social networks,the number of users of social networks is growing rapidly,and users' demand for personalized services is also increasing,which brings opportunities for the development of social networks:the ability to accurately capture user characteristics,build user profiles and provide targeted services has become the key for social networks to attract more users.Therefore,the user profiling method has been widely concerned by the academic and industrial circles in recent years.However,the number of users of the emerging social network platform is not yet large enough,and the data that can be directly used for user profile analysis is very rare and difficult to obtain.It is a great challenge to model user profile,analyze users and recommend information to users with only scarce data.If a well-developed social network can be used to help emerging social network construct user profiles,the problem of data scarcity can be solved better and the development of emerging social networks can be promoted better,that is,the method of user profile across social networks.In this thesis,a transfer learning based cross-network user profiling algorithm is proposed to help emerging social networks to build user profiles by using the information in the well-developed social networks,so as to solve the problem of insufficient user data and difficulty in obtaining user profiles in the emerging networks.This algorithm is composed of three parts.First,for the different characteristics of social networks,the characteristic-aware inner-domain attention layer is constructed,the user representation is obtained by combining the user's attribute information and structural information with the graph convolutional network,and the user representation is mapped to different representation subspaces through multi-head attention mechanism.Secondly,in order to capture the dependency information between users,the attention mechanism and similarity computing model are used to calculate the similarities between users.Thirdly,the problem of domain shift in the process of data transferring from the source domain to the target domain is solved through the adversarial learning method.The whole method framework achieves the final overall optimization by alternating iterative training.In addition,in order to achieve user profiling efficiently,this thesis proposes an improved algorithm for transfer learning based cross-network user profiling algorithm.Experimental results show that the cross-network transfer learning based user profiling method and the improved algorithm proposed in this thesis can effectively extract information from the social network of the source domain.So as to help the social network of the target domain to build user profiles and improve the accuracy of user profiles.