A New Collaborative Filtering Algorithm Based On Both Local And Global Similarity And Singular Value Decomposition(SVD)

Nowadays, with the Internet and the Internet of Things, which is gradually being established, it has been making it in the sea of information among the people for now.The "lack of information" era is gone, while information overload becomes a major problem to be solved.To solve the problem of information overload, the recommender systems come into being.To explore the Long Tail effect better, the collaborative filtering recommender systems are introduced. The collaborative filtering recommender systems are to fully study the user's personal interests, to locate the item for a certain specific user, and to provide users with a more personalized recommendation, thereby the collaborative filtering recommender systems put the accurate recommendations from the Nagao goods to users who need it, helping users find that the items who are interested in but it's hard for themselvs to find it. In a specific practical application scenario which uses the collaborative filtering algorithm, because in the beginning,the users, usually only give a small part of the evaluation or purchase a bite of goods, which will lead to an initial scoring matrix in the process of collecting the personal preferences of the user, the formation of( users- items) very sparse, which leads to collaborative filtering data sparseness and "cold start" problem. The challenge the collaborative filtering algorithm facing is to let the collaborative filtering recommender systems with relatively little effective rates to give more accurate predictions.For the ubiquitous data sparseness and "cold start" problem of the collaborative filtering rencommender systems, and to achieve the goal-- "effective with relatively few ratings to get an accurate prediction", we propose a new research on the collaborative filtering algorithm in the recommender systems- A new collaborative filtering algorithm based on both local and global similarity and singular value decomposition(L.G.SVD). The new algorithm is divided into three parts:(1) The scoring matrix(user- items) pre-treated by the similarity based on Local(Local User Similarity, L) and based on the Global similarity(Global User Similarity,G),which is collaborative filtering framework to process the initial scoring matrix(user- items) to obtain a new scoring matrix(user- items), as a next step algorithm inputs.(2) The scoring matrix(user- items) for SVD processing, using matrix decomposition model called singular value mathematical decomposition theory(Singular value decomposition, SVD) to re-process the scoring matrix(user- items)to obtain the active users neighbors by their predicted results.(3) The function of svd, using svd to obtain the final predictive value of the user and to give the final calculation recommendation. Public data set test results show that the algorithm, to some extent, alleviates the sparse and cold start problems, and improve the result of prediction accuracy and the recommendation accuracy.