User Online Behavior Vectorizing Model And Its Application

The further integration of the Internet and walks of life accelerates the speed at which the Internet penetrates into people's daily lives.The development of these applications and services,on the one hand,has accelerated the process of Internet inclusiveness;on the other hand,it has brought with the problems of “information-mazing” and “informationoverloading”,which makes it difficult for users to efficiently and conveniently find the content from the mass of information that really meets their desires.Fortunately,users had accumulated a large amount of personal behavior data in the process of participating in the creation,dissemination and consumption of Internet information,which provides an opportunity for the platform to model user behavior and render personalized,intelligent and customized services for users.Nowadays,vectorizing user preferences has been widely used to model user behavior.However,the explosive growth of user behavior data contributes to the suboptimality of the existing algorithm in modeling user behavior due to the issues about dealing with the massive,sparse,heterogeneous and low-quality data.Thus,this paper proposes three user online behavior vectorizing algorithms from the perspectives of bipartite network embedding and cross-domain transfer learning.Main contributions are summarized as follows:· User behavior vectorizing based on bipartite network embedding The interactive behavior between users and items can naturally form a bipartite network.Thus,we propose a bipartite network embedding algorithm,namely Bi NE,which can be used for user behavior vectorizing.The propose of the algorithm is to utilize the explicit and implicit relations mined from network topology structure to alleviate the problem of data sparsity.The model Bi NE designs variable-length adaptive random walk generator to capture the high-order implicit relationships between vertices and preserves the long-tail distribution of vertex degrees.In addition,the structure-aware negative sampling method greatly improves efficiency in optimizing the algorithm.· User behavior vectorizing based on co-factorizing multiple matrices To overcome the limitation of generating a large number of random paths when the Bi NE method captures the implicit relations between vertices,we derive the approximate equivalent matrix form of Bi NE.Thus,sufficient implicit relations between vertices in the network can be obtained without generating random paths.We also apply it to model user behavior via co-factorizing multiple matrices.· User behavior vectorizing based on transfer learning We enrich the information of target domain with the information of mature domains,and propose a user behavior vectorizing algorithm based on transfer learning,namely TLUM.The algorithm first takes the common users and items as the bridge to establish the relationships among domains.Then it uses similarity information between users or items to integrate the user behavior characteristics of mature domains into the target domain through the nearest neighbor-based migration method.Thus,TLUM can effectively alleviate the problem of data sparsity and low-quality.In addition,we establish the relationship between TLUM and vectorizing of vertices in the bipartite network.