Research On Many-objective Recommendation Model Based On Matrix Factorization

With the explosive growth of data volume,the problem of "information overload" has become increasingly obvious.It is difficult for the user to quickly obtain the required information from the massive amount of information,which causes inconvenience to the user's daily life.Therefore,simple information retrieval can no longer meet the needs of users,and recommendation systems have emerged.The recommendation system establishes a model,which accurately grasps the user's preferences and establishes the connection between the user and the item.As a result,it can recommend the relevant information and items needed by the user.Although the recommendation algorithm has effectively improved the efficiency of information retrieval and has been widely used in e-commerce,it still faces many problems.In this paper,the matrix factorization algorithm for novelty and diversity recommendation and the skewed matrix factorization algorithm for long-tail item recommendation are respectively proposed to make up for the single optimization objective and the neglect of long-tail items in the traditional recommendation algorithm.In addition,a many-objective two-layer recommendation model based on the DNMF algorithm is proposed.The main contributions of this article are as follows:(1)The traditional recommendation algorithm based on collaborative filtering only focuses on improving the accuracy,making the items in the recommendation list single and unable to attract users.In response to this problem,the DNMF recommendation algorithm is proposed.The algorithm includes two additional constraints: 1)Novelty constraints,so that the hidden factor vector of the target user is close to the average hidden factor vector of those users who have rated long-tail items.Improve the novelty of recommendation;2)Diversity constraints,make the latent factor vector of each item close to the mean value of the latent factor vectors of all items,thereby diversifying the recommendation list.This recommendation algorithm expands the loss function of traditional matrix factorization,while optimizing accuracy,novelty and diversity.Compared with the existing two-layer recommendation algorithm with inconsistent optimization objectives,the DNMF algorithm has a more comprehensive recommendation effect and is more excellent Performance.(2)Aiming at the problem of a large number of long-tail items,high profits,but a low recommendation,and unsatisfactory sales,the SMF algorithm is proposed.This algorithm uses two strategies to enhance the importance of long-tail items.1)The distribution of existing ratings is biased towards long-tail items;2)The weight of long-tail items is increased according to their low popularity.In addition,we also propose a new evaluation index of "Acceptable Novelty(hereinafter referred to as AN)" to evaluate the novelty of acceptable items(items that users like).Comprehensive experiments show that the proposed SMF algorithm can effectively improve the novelty of recommendations while ensuring the accuracy of recommendations.(3)Aiming at the low efficiency of a single objective recommendation algorithm in complex recommendation scenarios,a two-layer recommendation model MDRM is proposed that simultaneously optimizes the four recommendation objectives of novelty,diversity,accuracy and recall.The model consists of two improved algorithms: 1)bottom layer: DNMF algorithm is used to predict unknown item ratings;2)top layer: many-objective optimization algorithm generates recommendation lists for users.Compared with the existing two-layer recommendation model,the proposed model can effectively improve the four evaluation indicators of recommendation and has a more comprehensive recommendation performance.