Research On Collaborative Filtering Algorithm In Personalization Recommendation System

With the development of Internet and e-commerce, information overload is more serious and makes people to find information they need to spend more time and energy. Sometimes they may also be lost in the numerous information, forgetting what they really need. To an extent, the search engines can help the users to filter information, but this is only for those who know definitely what they need. For users with more vague requirements, it helps little. In this case, the personalized recommendation system came into being. It not only helps people to filter information or items, but also to recommend information or items which they are interested in. However, the current personalized recommendation system is also facing some challenges with the rapid increase in the number of users and the type of information.This paper focuses on the optimization and improvement of the cold start and the sparsity of the collaborative filtering algorithm in the current personalized recommendation system. First of all, the basic theoretical knowledge of the personalized recommendation system is introduced and sorted out in detail. Then, the basic idea and the common arithmetic of collaborative filtering algorithm is analyzed and summarized, and the shortcomings of the current collaborative filtering algorithm are sorted out, and then the reasons are analyzed. According to the results of the analysis, this paper proposes a collaborative filtering algorithm to add project properties category. This algorithm optimizes the traditional similarity measurement method, which increases a parameter of a project properties category when the similarity between items is calculated. This makes up for the deficiency that the traditional metrics compare between different categories of items when the similarity between items is calculated, which lead to the nearest neighbors of their projects not accurate. The basic idea of the algorithm is to classify items using project property, and then to use the improved formula to calculate the similarity in the class according to the classification of the situation, that is to add the formula of the project properties parameters to calculate the similarity between the projects. Secondly the nearest neighbor set of the target item is generated according to the calculation results of the similarity. Then its score is predicted in the class according to the nearest neighbor and put a higher score on the first N item as the Top-N output.Finally, using the open source data packet that the Movielens website provides to verify the improved algorithm, then the mean absolute error(MAE) is chosen as the standard to measure the accuracy of the algorithm. By comparing the MAE before and after the improved algorithm, it can be seen that the improved algorithm can reduce the influence of the cold start and the sparsity on the accuracy of the recommendation algorithm to a certain extent, and improve the effect of the recommendation system.