Research On Collaborative Filtering Recommendation Algorithm Faced To Sparsity Matrix Bias

Recommendation system rises in response to the proper time and conditionsto help users efficiently obtain the useful information they need from huge amounts of information.Recommendation system can learn the preferences of users through the analysis of the behavior data of users,and then take the initiative to recommend the interested information for users to meet their personalized recommendations.At present,nearest neighbor thought,the similarity thought,weighted thought are still the most commonly used in all kinds of recommendation algorithm thoughts,but as the expansion of the number of users and the scale of systems,recommendation system is faced with the problems of data sparsity,popular bias,poor extendibility and so on.Traditional recommendation algorithms were influenced by data sparsity,the accuracy of similarity calculations were insufficient,which leads to the research of nearest neighbor is not accurate.According to the problem of data sparsity,this paper puts forward the collaborative filtering recommendation algorithm based on improving Singular Value Decomposition,the similarity between users are no longer calculate with score matrix,but through the user eigenvector matrix.Its concrete realization method is through the decomposition of singular value to get users' and items' eigenvector matrix to calculate the corresponding characteristics of similarity,and take advantage of the potential relationship between users and items,using singular value to extract some essential features to calculate the similarity of the corresponding eigenvectors of two users,then obtain the similarity between users and items.In addition,traditional recommendation algorithms were tend to recommend products with higher popular degree.According to popular bias,this paper introduce the penalty function into traditional algorithm according to the item popularity and the user interest information,then put forwards the collaborative filtering recommendation algorithm based on the degree of punishment.This paper analyzes from the effect of the produce process of user behavior data on user model,build punishment functions according to items popularity and user interest degree and use the penalty function to adjust the weight of different popularity items in user model.On the MovieLens100 K data set,it verified the performance of the traditional recommendation algorithms faced with sparse matrix bias,the experimental results show that the accuracy of the collaborative filtering recommendation algorithm based on SVD similarity have increased by 0.68%,improving the accuracy of nearest neighbor search,and it is suitable for news recommend areas.The fraction of coverage in collaborative filtering recommendation algorithm based on penalty is 1.47% higher than the user based,alleviating the impact of popularity on user behavior,it is suitable for electronic commerce recommendation.