Construction And Optimization Of Virtual Information Core In Collaborative Filtering Recommendation System

The recommendation system is an information filtering system.By analyzing users' historical behavior data,it helps users filter out spam in massive data and recommend appropriate items for users.Collaborative filtering algorithm is one of the most widely used algorithms in the recommendation system,which predicts the unknown ratings of items by analyzing the preference information of neighbor users.However,as the size of the data increases,the algorithm also exposes some typical problems,such as data sparsity and scalability issues.In order to alleviate the above problems,the information-core-based collaborative filtering is proposed,where information core is a set of real users with more useful information by eliminating some potential noisy users.However,the recommendation accuracy obtained by the information core composed of real users is usually not high enough due to some information loss.To overcome this shortcoming,this thesis proposes several recommendation methods based on virtual information core composed of virtual users,which further alleviate the scalability issues while obtaining higher recommendation accuracy.The improvements mainly start from two aspects: the first is to construct the virtual users with richer recommendation information;the second is to select some virtual users to form the virtual information core to achieve better recommendations with fewer users.The main work of this thesis is summarized as follows:(1)A method of virtual information core optimization based on clustering and evolutionary algorithms is proposed.The rating matrix is first reduced in dimension by t-SNE algorithm to obtain the users' low-dimensional matrix.Then,in order to make the users' information fully utilized,users are clustered repeatedly based on the low-dimensional matrix.Finally,the cluster centers are used to form the virtual user set,from which the virtual information core is selected based on evolutionary algorithm.When evaluating the fitness,a smaller training set and validation set are constructed to reduce the fitness evaluation cost and accelerate the efficiency of the algorithm.The experimental results illustrate that this proposed method has excellent recommendation accuracy.(2)A method of constructing the virtual information core based on clustering and multi-armed bandit is proposed.This method is an improvement of the first method and its main motivation is to improve the efficiency of virtual information core selection.Firstly,based on the virtual user set that has been obtained,the multi-armed bandit is used to select the virtual information core.In this step,the process of exploitation and exploration of the multiarmed bandit is realized by the ?-greedy algorithm,and the reward mechanism is redefined to reduce the influence of noise feedback.Then,the reward information of each arm is updated based on the ratings bias,thereby obtaining the trust value of virtual users.Finally,the virtual information core is composed of some virtual users with high trust value.The experimental results illustrate that the recommendation accuracy of the virtual information core selected by this method is higher,and the selection efficiency is obviously improved.(3)A recommendation method based on user similarity and multi-arm bandit is proposed.This method aims to improve the construction time and recommendation accuracy in the second method.Firstly,the multi-arm bandit is used to select some users from the real user set as the initial users who are regarded as the initial cluster centers.Secondly,based on the similarity with the initial users,the remaining users are repeatedly clustered to several similar clusters.Thirdly,the cluster center of each cluster is updated and the virtual user set is built using the new cluster centers.Finally,the multi-arm bandit is used again to select virtual users to form the virtual information core.The experimental results illustrate that this method greatly shortens the time of constructing virtual users and improves the efficiency of the algorithm.In addition,the recommendation accuracy of the virtual users is satisfactory.