Research On Personal Movie Recommender Algorithm Based On Matrix Factorization

In the social context of big data,the amount and type of information is growing rapidly.We need to spend a lot of time searching for interesting or useful information,and information overload has occurred.With the advent of personalized recommendation systems,the problem of informat ion overload has been alleviated.Recently,a large number of related researches on recommendation systems have emerged.Due to the advantages of high accuracy,high scalability,and fast operation of matrix factorization,it has quickly become a research hotspot in the field of recommendation algorithms.In this paper,the traditional similarity calculation is improved based on the principle of hybrid recommendation algorithm,and the probability matrix decomposition algorithm(PMF)is improved based on the perspective of distance,and the bayesian personalized ranking algorithm(BPR)is improved based on the idea of neighbor recommendation.This paper uses the interest drift factor and hot item factor to improve Pearson similarity by weighting.The singular value decomposition algorithm(SVD)is used to obtain the user's implicit feature matrix,and the user's implicit feature similarity is calculated based on the cosine similarity.Finally,the rating distribution of users is analyzed using Hellinger Distance,which is used as the weight of Pearson similarity that combines with the similarity of implicit features.Through experiments,it is compared with the traditional Pearson similarity,Jaccard similarity,and JMSD algorithm.The results show that the MAE of the CFHPI algorithm reaches 0.769,which is at least 0.015 lower than the other three algorithms,and Recall reaches 0.187,which is at least 0.01 higher than the other three algorithms.It is proved that CFHPI algorithm does improve the accuracy of user similarity.This paper predicts the score from the perspective of distance,and measure s the user's preference for movies by the distance between the user feature vector and the item feature vector and based on the idea of the neighbor recommendation algorithm,the trust of user ratings is analyzed,and the global trust is calculated for each rating.Therefore,this paper proposes a probability matrix factorization algorithm based on distance and trust(DTPMF),through experiments,DTPMF is compared with the SVD,PMF,and bias PMF algorithms in two datasets.DTPMF reached 0.915 and 0.867 at RMSE,compared with the other three algorithms,it is reduced at least 0.013 on Movielens?100k and at least 0.022 on Movielens?1M.It is proved that the superiority of DTPMF in improving the accuracy of rating prediction.In order to express the user's preference for the item more acc urately,this paper builds a user's neighbor matrix based on the idea of neighbor recommendation algorithm.And combined with the original rating matrix,the user's preferences for the project are divided into three categories: high preference,medium preference and low preference.According to the divided three types of preferences,a triple is constructed for the BPR algorithm and then optimized,this paper proposes an improved Bayesian personalized ranking algorithm based on nearest neighbors(NBPR).Through experiments,NBPR is compared with the BPR and pop Rank algorithms on two datasets.The NBPR algorithm's Recall values are 0.087 and 0.065,respectively.Compared with the other two algorithms,it is improved at least 0.017 and 0.013 on the two data sets,respectively.Precision values are 0.41 and 0.46 respectively.Compared with the other two algorithms,it is improved at least 0.017 and 0.012 on the two data sets,respectively.It is proved that the advantages of NBPR algorithm in improving the accuracy of ranking prediction.