Research On Recommendation Technology Based On Collaborative Filtering

Collaborative filtering algorithm tried to calculate their relationship between the users or the projects by analyzing user’s history behavior and records in order to determine each user’s nearest neighbor set or the nearest neighbor set of each project. Then, according to the characteristics of the set, the system recommends projects to the target user which he will be interested in. But the technology still faced with some problems, such as data sparsity, cold-starting, systemic vulnerability and other issues. This paper has been studied in-depth of the sparseness problem and improved the similarity calculation method, and this paper proposed an algorithm based on direct and indirect evaluation score for collaborative filtering in this paper, which has higher recommendation accuracy in the case of sparse data, and improved the quality of recommendation system.Firstly, we presented a calculation method based on the weighted similarity. Due to the data sparseness problem the users’common rating-items are very less. In this case, the traditional similarity calculation method has a certain chance problem in the calculate of similarity result which had a bad influence on the selection of the nearest neighbor set and reducing the accuracy of the system, this algorithm is mainly to solve the accidental problem which existing in the result of similarity calculation, to improve the quality of recommendation.Secondly, a collaborative filtering algorithm based on both direct rating and indirect rating is proposed. For direct rating, weighted user similarity and weighted item similarity are defined, direct recommendation user set and direct recommendation item set are constructed, and computation method of direct rating weight is presented. For indirect rating, similar rating set is constructed, and rating similarity is defined. Furthermore, comprehensive rating weight is defined, which is used for producing the final recommendation results. After a large number of experiments, we found that the proposed algorithm can achieve high quality on datasets with different sparse levels.Finally, this paper realized the proposed algorithm as a prototype system, and analyzed the quality of the recommendation system.