Research On Personalized Recommendation System Algorithm Based On Matrix Factorization Technology

In the era of DT,a large number of various items of information flowed into the Internet,resulting in the "overload" of network information.The overloaded item information does not stimulate the user's consumption excessively,but instead causes the user to "choose difficult" problems.The Personalized Recommendation System came into being in this context.The Personalized Recommendation System hopes to provide personalized and accurate recommendations for various consumer user groups.The rating prediction recommendation algorithm is the cornerstone of the recommendation system to complete the personalized recommendation task.In recent years,it has received extensive attention and research from the academic community.Unfortunately,studies on the problems of rating prediction recommendation algorithms have been mixed.Some studies have focused on improving the rating prediction accuracy of cold users(users with no history rating data in the system),and some studies have focused on improving the rating prediction accuracy of warm users,and some studies have focused on the accuracy of rating predictions for cold-start items.That is to say,there is no universal rating prediction model,starting from both the user and the item,to improve the rating accuracy of cold and warm users simultaneously.Therefore,this paper proposes a general model based on the Bayesian Probability Matrix Factorization(BPMF)model.This model integrates explicit and implicit relationship network information of users and items to assist in feature extraction of them,thereby improving the rating of cold and warm users.Specifically,this article did the following work:(1)According to whether the relationship network between users(items)comes from rating data,the relationship network extracted from data other than rating data is called expicit relationship network of users(items).On the contrary,the relationship network extracted from rating data is called implicit relationship network of users(items).It is believed that explicit relationship network can assist cold users(items)in feature extraction,while implicit relationship network can do a good job of modifying the feature of warm users(items).(2)Successfully merged user(item)explicit and implicit network with the BPMF model,and presented a probabilistic reasoning graph of fusion model(BPMFG),and introduced the extraction method and correction method of user(item)features in detail.In the aspect of feature vector correction,different correction coefficients are introduced for cold and warm users(items)to ensure that the correction of feature vectors is not distorted.In the details of the model implementation,different mean and different variance characteristics of the hyperparameter distribution of the user(item)are considered,and the user rating bias and item scoring bias are considered,too.(3)Based on the PageRank algorithm,a new user trust degree calculation method is proposed for the existing problems in the user traditional trust relationship calculation in the trust network(a user explicit network).Further,for users with less trusted users,a method for calculating the trustworthiness of k-hop-trusted users is proposed.(4)Comparison of experiments on three real rating data sets shows that the algorithm is superior to the four performance rating prediction algorithms,the BPMFG model can effectively improve the rating prediction accuracy of cold and warm users under the condition of uneven distribution of rating data.Compared with the model with the best performance,the Root Mean Square Error(RMSE)of the BPMFG model's rating prediction for cold and warm users improved by an average of 6.04% and 5.18%.(5)In addition,in order to verify the extended performance of the BPMFG model,and also provide the industry with a distributed implementation idea of the Probability Matrix Factorization series recommendation model,this paper parallelizes the BPMFG model based on Spark-GraphX.A cluster with 6 nodes was built and the distributed training performance of the model was tested through three groups of experiments.