Research On The Recommendation Algorithm Of Tagging Collaborative Filtering And Some Improved Strategies

The development of new Internet technology has led to the emergence of new applications.Social network and e-commerce are typical new network applications.The development of this new application has also promoted the birth of new technology.As the amount of information increases,both e-commerce and social networks need a system to discover the commercial value of a large amount of information.The recommendation system is this technology.However,the recommendation system also faces problems such as cold start,data sparsity,large amount of computation,and uncertainty of prediction accuracy.In this paper,the research on the recommendation system is mainly to provide some new ideas for solving these problems.There are many ways to solve the problem of data sparsity.You can solve the problem with methods or ideas like basic information of users,the therapy of two maps or the default voting method.You can also use the data "smooth" technique to solve the sparsity problem.The cold start problem is a special data sparsity problem.In addition to the above methods,some hybrid algorithms can also be used to solve this problem.In order to solve the problem of large quantity of data and low quality of prediction results,This paper carefully studies the advantages and disadvantages of the traditional recommendation algorithm,and finds some improvements.Combined with statistical thinking,clustering thought and modeling idea,it puts forward three improvement strategies for label system and collaborative filtering recommendation algorithm.First,on the basic of traditional collaborative filtering,the preference matrix of users and items is obtained by computing the rating matrix of users and items.By using the idea of TFIDF,we calculate the user preference matrix of tags and the correlation matrix between labels and items,then calculate the user's preferences for goods so as to complete the recommendation.Secondly,in order to solve the problem of large amount of data,this paper proposes a method of spectral clustering for effective data filtering.Based on the above method,using the idea of clustering using spectral clustering method,users will be divided into several categories,according to the number of users in the class,greatly reduced the amount of data of the pressure.Finally,in order to fully tap the tag information,we use LDA algorithm to tap the theme from tags.What's more,we will get the user-tag correlation degree,item-tag correlation degree and user-item correlation degree through calculating relative data with the combination of above two methods.This paper carries out experiments to verify the three methods with real movielens datasets.Experiment results show that compared with the basic collaborative filtering,these three algorithms have better performance in recall and accuracy.