Research On Collaborative Filtering Algorithm In Recommendation System

The rapid development of information technology speeding up the creation and propagation velocity of information, so that knowledge can be more freely sharing and communication. But massive information also brings negative effect which cannot be ignored- information overload. Information retrieval and filtering technology such as search engine to some extent alleviate the information overload problem, but these techniques require user to have clear goals, and cannot intelligently screen objects which user may be interest. Recommendation system then born and widely used. One of the most popular algorithm is collaborative filtering algorithm which has been made a great deal of attention and research, and achieved good effect. It has important research value and application prospect.High accuracy of recommendation is the main goal to recommendation system. This thesis proposes a new similarity calculation method which starts from user rating data, and classifies user common rating items into two class, one is the same rating items and the other one is the different rating items. Then the similarity could be calculated from this two types items based on the rating between users. This method is called the rating similarity. Finally, improving the original algorithm by fusion rating similarity.At the same time, sparsity problem is also a key factor that cannot be avoided. The exponential growth of object and user not only affects the recommendation system efficiency, but also produces a high dimensional data grid which is very sparse, causing the historical data available limited. So as to reduce the effect, this thesis proposes an improved collaborative filtering algorithm. It first calculates the missing data based on the rating similarity collaborative filtering algorithm and Slope One algorithm, and use it fulfill the sparsity matrix. Then use the collaborative filtering algorithm to generate recommendation based on the filled matrix. Finally improve the system performance in sparse environment.At the end, verified the proposed two algorithms by experiments.