Research On Algorithm Of Personalized Recommendation In Digital Library

Recently with the improvement and popularization of digital library, the resource and information of digital library are getting more and more abundant. However, while we share enjoy the convenience and fast speed of digital library, we are puzzled in front of such a mass of information. In a word, digital library with mass information is not able to fast, effectively and intelligently provide user with the information which he is interested in.The individuation recommendation service of digital library is user centered, analyzing the habit, behavior, occupation, and custom of the user, and then provide the user with the information his is interested in intelligently and effectively. Collaborative Filtering recommendation algorithm is the one of most used algorithms and we can explain its core idea in this way: if the users, who have the same interests with me, like this product, I probably like it as well.Collaborative Filtering recommendation algorithm recommends a certain item for the user out of the rating score history of the user. In a real system, the data of users and items is too huge, while the number of items for the user to rate is comparatively small, only nearly 1% of the whole items in the system. So the problem of data sparsity of user rating score is apparent which impairs the accuracy of the final recommendation result. This is the so-called problem of data sparsity, which exists ubiquitously in a real digital library.This dissertation aims at how to improving the accuracy the recommendation result in data sparsity environment and make it better to be applied to the individuation recommendation service of digital library. The main research content could be described as"tow algorithms and a system".1. Proposed one similarity algorithm based on the nearest neighbor dynamic reordering. Taking full use of categories information with sparse data and dynamically adjusting weights of users in neighbor sets according to different project targets will describe the similarity of users more accurately. What's more, proposing one overlap factor to make up the shortage of the existing methods to manually adjust the parameters in order to enhance the usefulness of the algorithm. Experiments show this algorithm will improve the recommendation accuracy much with sparse data. 2. Proposed one trust-based collaborative filtering which builds relations between users without any common rating items to solve difficulty to find nearest neighbor sets caused by data sparsity. And Experiments show it would relieve difficulty to forming effective neighbor sets by defining trust degree and trust-based disseminating principles to extend traditional nearest neighbor sets.3. We apply the proposed algorithm to the library interaction system for education and research to have a test on its practicability.Finally, we draw conclusions and discuss the future work.