The Recommendation System Based On Collaborative Filtering

With the rapid development of Internet, the information overload is becoming serious. The trend for the next development of Internet technology is to solve the information overload problem, provides the interested information based on the difference of each user. Therefore, the personalized recommendation technology emerged, which is based on the difference between the users and uses the corresponding algorithm to help user find their favorite content. In recent years, among the academic, the research heat on recommendation system is increasingly high. The recommendation system has gradually formed an independent discipline. The major Internet companies also have strong investment on their own recommendation system. But, the cold start, data sparsity, scalability problems of recommendation system are still not been fully resolved and the recommendation attack methods against recommendation system is also increasing. In this paper, the research focus is personalized news recommendation system with a high efficiency and scalability which is based on collaborative filtering. The main works are:1) Propose efficient clustering and similarity calculation method. Calculate the recommendation score based on the collaborative filtering which also combines the accompanied views. Solve the problem of sparse matrices and the cold start of users.2) For the characters of news and information, classify the articles according to the news topic. Predict the current interests of users based on the history behavior to create user configuration file, which is used to filtering the recommendation results.3) Implement the scalable offline calculation algorithm based on MapReduce model, which will lead to the personalized information recommendation system running in parallel to meet the personalized recommendation requirements of massive information and massive users.4) Provide an overall design of general personalized recommendation system, including online part, offline part and their workflows. According to the characteristics of news area, provide the design of prediction module for user interest themes.Finally, the above algorithm was validated through efficiency MinHash recommended after clustering increases the amount of data and more obvious advantages; based MinHash and frequent queues hybrid recommendation algorithm, is recommended for cold-start user-generated precision and recall relative accuracy rate with traditional recommendation algorithms and recall rates over50%increase; based MinHash clustering, frequent queues, topics of interest to predict hybrid recommendation algorithm, the results generated recommendation precision and recall rate compared with the traditional recommendation algorithm precision and recall rate of more than10%increase.