Improved Collaborative Filtering Recommendation Algorithm Based On BP Neural Network

With the rapid development of mobile Internet and the technology of big data,the problem of information overloading is increasingly prominent.In order to improve the efficiency of getting information from mass data for users and the enterprise production profit,recommendation algorithms have been widely used in different applications.Collaborative filtering(CF)is the most widely used personalized recommendation algorithm among the commercial recommendation systems,while measuring the interest similarity among users is the most crucial step of this algorithm.With the increase of the number of users and items,the computation cost of the similarity matrix of users also increases,thus the traditional CF algorithm has limitation in extensibility.Moreover,the traditional CF algorithm measures the interest similarity among users by calculating the similarity of their rating vectors.Under a highly sparse rating matrix,its performance is not satisfactory and the rating prediction accuracy should be improved.Targeting at the above two problems,we consider the influence of the attribute features of items on the users' rating behavior,set up a user rating prediction model regarding item attribute features using the BP neural network(BPNN),and improve the traditional CF algorithm based on this model.The contributions of this thesis are as follows:First,regarding the extensibility limitation of the traditional CF algorithm,we propose an improved collaborative filtering recommendation algorithm based on clustering analysis on user rating preference features.The new algorithm uses the rating prediction model to extract the low dimensional rating preference feature vector of item attribute features of users.Then we cluster users by analyzing those feature vectors so as to reduce the time complexity and the cost of memory space of the traditional CF algorithm.Secondly,regarding the low prediction accuracy of the traditional algorithm under sparse rating matrices,we propose an improved algorithm based on the prediction error of neighbor users.This algorithm measures the similarity of interest preference among users from different aspects.The candidate neighbor user takes the history rating records of the target user as the prediction projects,and then uses BPNN rating prediction model to predict those projects.The smaller the prediction error,the more similar the rating behavior between the candidate neighbor user and the target users.Based on Movielens open source data set,we test the proposed two algorithms and compare them with some existing algorithms.The experimental results show that compared with other algorithms the two proposed algorithms have better performance in extensibility and accuracy,respectively.