Research On Recommendation Algorithm For Group-Users' Temporal Behaviors

The amount of online data is increasing exponentially with the so rapid development of Internet technology,and it is becoming increasingly hard for users to find accurate content from massive amounts of data.Then the emergence of recommendation systems provides ideas for solving this contradiction.By using the user's interactive behavior on information,the recommendation systems connects users with the information they are interested so that helping users discover their potential interests from massive information.How to utilize the users' historical behavior data to predict users' preferences and discover users' potential needs,and then actively recommend personalized content to users has become a hot issue of continuing concerned in recommendation filed.Existing recommendation methods can achieve good recom mendation effects in the scenarios of simple user components and user's interests do not change over time.However,the temporal recommendation for group-users is still a challenge nowadays.The main difficulties are:(1)User behavior data is no longer generated by a single user.For example,for the TV program on-demand scenario in a real-world family,the user's viewing records generated under a single user ID imply the watching behaviors of multiple family members,and each member's preferred program type is different from each other;(2)Users' interests will drift with time.And for family members,their preferences change with time in a different law.In response to the above difficulties,this thesis introduces a time-varying group role multinomial matrix and proposes Dynamic Recommendation algorithm for Group-users' temporal behaviors called DRGu,which based on the idea of matrix factorization.Its main work is: First,divide group roles to address the hidden group problem in user behaviors,and introduce a multinomial distribution matrix of group roles that changes with time according to the user's behavior data to infer the group role type generated behaviors at the moment;On this basis,the temporal characteristics and group-role behavior characteristics are integrated into the matrix factorization model forming a time-varying group-user preference model to capture the dynamic preferences of group-users;At the same time,the problem of items' long-tail in time slices is more prominent during the temporal behavioral learning,so we also construct a dynamic exposure model based on the popularity calculation by introducing time-varying exposure variables,which converts the problem into a Inverse Propensity Weighted method called IPW for the training samples to improve the effect of temporal recommendation for group-users.The results of experiments on two real datasets,the Internet online TV on-demand program IPTV dataset and the cloud theme scene shopping dataset Cloud Theme in Taobao APP,show that compared with the existing correlation matrix decomposition method s,the temporal recommendation algorithm proposed in this thesis can effectively make dynamic recommendations for group-users.