Collaborative Filtering Based Content Recommendation In Broadcast-storage Systems

As the Internet traffic increases,it is difficult for existing Internet architecture and content transmission technology to assure efficient information sharing.Broadcast storage systems(BSS)bring broadcast distribution into TCP/IP.Information resources of common interest are distributed to edge servers near users via radiation distribution to build an efficient way of information sharing in BSS.BSS have clear advantages in reducing the redundant traffic in the Internet and remitting information overload problem.Uniform content label(UCL)is the basis of BSS.Users read UCL to determine if it is necessary to request the full text.However,due to the huge amount,users can easily get lost among them.How to recommend personalized UCL to users and realize the interest matching become the keys of BSS.Collaborative filtering(CF)is a widely used method for interest matching.However,existing researches show great limitations in BSS.Firstly,in similar neighbors discovering stage,traditional methods perform measurement based on the numerical ratings of users on UCL and overlook the abundant attributes of UCL and users.Secondly,in interest prediction stage,existing methods always have parameters independent on dataset and once the dataset changes,the prediction performance gets unstable.Thirdly,in recommended results generation stage,traditional methods need to train parameter to control the diversity level that is not suitable for dynamic datasets.Besides that,they do not take into consideration both the ratings and the semantic information of UCL and converge slowly thus resulting in low performance of UCL recommendation and restricting the performance of BSS.Aiming at solving the above problems,this dissertation studies the CF based content recommendation technology in BSS from four aspects.Firstly,we study the efficient users clustering and UCL classification algorithms and then propose similar neighbors discovering methods considering user social information and UCL attributes to lay the foundation of follow-up interest prediction.Secondly,we study the interest prediction methods driven by the data density feature to enhance the prediction stability when the dataset changes dramatically.Thirdly,we design a data structure called timing sensitive semantic cover tree and propose a non-weighted-parameter result diversification method based on it to ensure the UCL recommendation accurate and round in dynamic BSS.We also enhance the novelty and response speed of the diversification process.Finally,we design and implement the UCL recommendation system in real BSS environment.We also conduct comprehensive function test of its modules to verify the effectivity and feasibility of the research of this dissertation.The research of UCL recommendation in BSS is studied in this dissertation.Results from lots of simulations and experiments in BSS show that the algorithms in this dissertation can recommend UCL efficiently and improve the performance of BSS.The theoretical researches also give a significant value to the personalized recommendation in other realms.