Research On Personalized Recommendation Algorithms Based On Collaborative Filtering Technology

With the rapid development of Internet technology, information on the Internet is growing exponentially, which brings us both benefits and disadvantages. Internet has brought great convenience to our daily life and diversified information needs. However, it also has caused many problems, such as ‘Information Overload’, ‘Information Trek’, and so on. Personalized recommendation can solve those problems effectively. Collaborative filtering(CF) technology makes use of the relationship between users and items to create recommendation for the target user. It can deal with unstructured data very well. During the past two decades, CF technology has attracted much attention and become an important and hot problem in both academia and industry. Although CF technology has been used in e-commerce website widely, the traditional CF algorithms are still faced with sparse data, performance and scalability issues.This paper focuses on the sparse data, performance and scalability issues in the personalized recommendation algorithms. We mainly study the similarity measure in CF algorithms, and combine matrix factorization algorithm with user based collaborative filtering. In this paper, two classical algorithms are improved. The main works of this paper are listed as follows:1) The similarity in traditional CF algorithm is not accurate in sparse data. We combine the structural similarity with the traditional similarity and propose the combined similarity which can compensate the lack of accuracy in sparse data. As a result, based on combined similarity measure, an optimized collaborative filtering algorithm is proposed.2) Based on basic matrix factorization and biased matrix factorization, we put the user’s neighborhoods model of collaborative filtering algorithm into matrix factorization model. Combined with the new similarity in this paper, a matrix factorization algorithm based on user’s neighborhoods constraint is proposed.3) Finally, in the open Movie Lens data set, we perform several experiments on the improved algorithms. Experiment results show that the improved algorithms which proposed in this paper have better precision, recall and coverage. To a certain extent, the algorithms improve the quality of recommendation.