Research Of Recommendation Algorithms Based On Matrix Factorization

With the rapid development of e-commerce,the number of information commodity websites has increased dramatically.Users often need to spend a lot of time screening products they like or information they are interested in.The recommendation system collects users' historical behavior on the website,such as their shopping behavior on Taobao website,praises a piece of information on Weibo,and then trains the recommendation algorithm to get a list of recommended goods or information for specific users to meet users' personalized needs.Personalized recommendation system is not simply calculating hot commodities or hot information,but pushing different results according to different interests of each user.Personalized recommendation algorithm is a hot research issue in both academia and industry.Recommendation algorithm can be divided into score prediction algorithm and item ranking algorithm.In the problem of score prediction,general users have explicit feedback behavior,such as users scoring favorite movies high on Douban Film Scoring Website,and disliked movies low on Douban Film Scoring Website.In item ranking problem,users click on a news without scoring or praising.There are only positive samples in the system,while the other samples are a mixture of negative samples and missing values.Traditional recommendation algorithms include memory-based collaborative filtering algorithm,content-based algorithm and hybrid recommendation algorithm.Collaborative filtering algorithm is easy to understand,easy to operate and can achieve personalized recommendation,but it is vulnerable to the impact of data sparsity.Content-based algorithm only pays attention to the attributes of a user-interested project,which can solve the problem of cold start of the project.However,the recommendation result is single,and it is difficult to find potential products of users' needs.Hybrid recommendation algorithm combines the advantages of multiple algorithms to deal with complex recommendation scenarios,and the recommendation result is the sum of the recommendation results of multiple algorithms.These algorithms are vulnerable to data sparsity,so some scholars put forward matrix decomposition model,which has better performance in sparse data sets.Updating parameters with machine learning method can quantify user taste and commodity attributes more accurately,which greatly improves the accuracy of recommendation algorithm.Using scoring matrix as data source can no longer meet the requirement of recommendation accuracy.Some scholars excavate the global project similarity relationship.Some researchers propose that using heterogeneous auxiliary information,such as social network relationship,the relationship between users and others on social network can help the model capture users' needs and preferences.Collaborative filtering is a classical recommendation algorithm for fractional prediction,but it performs poorly on sparse data sets because similarities between many commodities in sparse data sets cannot be calculated.In this paper,fiss algorithm is proposed by utilizing social network assistant information and mining the similarity of global items(commodities,information,etc.)and analyzing the similarity and trust reliability between friends of social network.Its essence is to use social network relationship information to better calculate user feature vectors.Experiments on filmtrust movie data set and Douban data set show that fiss algorithm has less error than other social network algorithms.Pairwise method has the highest accuracy in relative ranking problem.It uses two low rank implicit eigenvalue matrices and sigmoid function to get users' personal preferences.The model assumes that the individual preference of the products that users have browsed is larger than that of the products that they have not browsed,and the ranking probability of the items is obtained.The minimum extremum objective function,i.e.loss function,is obtained from the maximum posteriori probability estimation of all data sets.In this paper,an algorithm FSBPR based on global similarity of items and Bayesian personalized ranking framework is proposed.FSBPR gives different weights to positive and negative samples to improve the accuracy of recommendation.Experiments on filmtrust and ml100 k datasets show that the recall rate of FSBPR algorithm is higher than other item ranking recommendation algorithms.