Research On Recommender Algorithms Based On Collaborative Filtering

With the continuous development of the Internet,the data source has expanded rapidly.However,because of losting in the ocean of data,it is difficult for users to find the information they need accurately,which makes the problem of information overload highlight.With the background,Recommendation System comes into being and the collaborative filtering recommendation algorithm is the most popular recommendation algorithm in the field of recommendation system.With the deepening of the research,it has been made great progress,but there are still some problems:(1)The sparsity problem leads to a serious decline in the accuracy of recommendation;(2)The top-N method has low recommended precision;(3)The difference of evaluating criteria of users causes that the user's ratings couldn't reflect the user's preference reasonably which affects the accuracy of the similarity calculation;(4)The problems of the popular items ratings interfering in similarity calculation;(5)The one sidedness of the recommendation group selection leads to the unreasonable nearest neighbor selection.This paper proposes some solutions to solve above problems.The main work and innovation points are as follows:(1)In order to solve the problem of sparsity and the low recommended precision of Top-N method,this paper proposed a collaborative filtering recommendation algorithm based on rating matrix filling and item predictability.Firstly,the matrix filling algorithm was used to improve the density of rating matrix;Then the credibility of the filling ratings was measured by the confidence coefficient C(the credibility of the true ratings is 1);Finally,the concept of item predictability was proposed which was used with the item's predictive ratings comprehensively to select the optimization recommended list.We proposed to formulate the problem similar to a special class of 0-1 knapsack problem.The experimental results show that,the algorithm can alleviate the influence of sparsity and improve the recommendation accuracy.(2)Among traditional collaborative filtering recommendation algorithms,the difference of evaluating criteria of users caused that the user's ratings couldn't reflect the user's preference reasonably.In order to solve this problem,this paper proposed the concept--Satisfactory Intervals(SI),and a collaborative filtering algorithm.Firstly,the algorithm established the relationship between users' ratings and SI.The algorithm solved the problem of evaluating criteria by partitioning SI,which could be more reasonable for users to express their preference.Then it calculated the similarity between users through the SI.Finally,this algorithm rated the item by its satisfaction which was calculated before.Experimental results show that this algorithm can solve the problem effectively and achieve better accuracy of recommendation obviously.(3)In order to solve the problem of the popular items ratings interfering in similarity calcu-lation this paper proposed the ratings distribution recommendation model.Based on this model,we designed a new collaborative filtering algorithm.According to ratings distribution,this algorithm firstly get the amount of information carried(The Shannon Entropy).Then,it calculated the rating weights to filter into traditional similarity calculation.The experimental results show that the algorithm can effectively alleviate the above problem and improve the accuracy of recommendation of the algorithm.(4)The one sidedness of the recommendation group selection leads to the unreasonable nearest neighbor selection.In order to solve the problem,a collaborative filtering algorithm based on user interests is proposed.Firstly,the algorithm calculates the category similarity between items,and clusters the items by the similarity;Then,it determines the interests of user according to the clustering results,and the penalty function is introduced to alleviate the impact of items popularity;Finally,it selects recommendation group and nearest neighbor sets dynamically according to user interests.The experimental results show that the proposed algorithm can effectively improve the rationality of the nearest neighbor selection and improve the recommendation accuracy.