Research On Collaborative Filtering Recommendation Algorithm For The Digital Library

As the digital library has a wealth of information resources and convenientoperability, the speed of its popularity and development is very quickly in recentyears. The more successful examples of current digital library are the ACM DigitalLibrary, China National Knowledge Infrastructure(CNKI) and so on, however, all ofthem face the problem that users can not quickly find the resources they real need,personalized recommendation technology is the current best way to solve thisproblem. The application of personalized recommendation service in the digitallibrary, that not only provides users with a personalized service, simplifies users’operation, but also improves the utilization efficiency of digital library resource, thetwo sides reached win. Recommended technology is the main contents of my thesis.Among the many recommended techniques, collaborative filtering recommendedtechnology is one of the technologies which is widely used in personalizedrecommended technology, and the thesis focuses on the user-based collaborativefiltering recommendation technology. The collaborative filtering technology facesmany problems, and the data sparsity problem is most severe. This problem causesthat the traditional algorithm can not accurately calculate the similarity between users,thereby affecting the final recommendation results. Although the traditional methodwhich fills scoring matrix with the default can alleviate this problem, but theconsequent is not satisfied. Data sparsity problem is also one of the main researchcontents of the thesis.The main contents of this thesis includes: for the data sparsity problem, the thesisputs forward the improved item rating prediction method, that is merging items whichusers rated as a common rating items between users, and using the similarity betweenitems to predict the unscored-item rating, the method can accurately calculate thesimilarity between users; Meanwhile, the thesis has improved traditional Pearsoncorrelation coefficient formula, and propose the concept of evaluation factor, in orderto more accurately measure the true extent of the similarity between different popularusers; Finally, the thesis has improved the prediction score formula, through calculating the similarity between items by label-based method and item-basedcollaborative filtering method, to predict the target item score which neighbor usershave not rated.On the basis of the movieLens dataset, the thesis verified that the improvedalgorithm has a higher precision than conventional recommended algorithm, theimproved one can effectively measure the similarity between users in the data sparsitysituation. Finally, we applied the improved algorithm to a data libraryrecommendation system, and got the corresponding recommended results, whichshowed that the improved algorithm has better recommendation accuracy thantraditional algorithm.