A Collaborative Filtering Of Local Optimization Selection For Recommendation Algorithm

The popularization of Internet and the fast development of information technology lead us to the times of information explosion, which on one side satisfies users’need for various information. But on the other side, as the Internet resources grow explosively, users have difficulty in finding the required ones when facing diverse information. And this is the problem of information overload. So personalized recommendation technology is proposed accordingly, which offers different users a more convenient way to acquire information. The technology itself recommends the information that customers are interested in based on their interests and relevant habit records.At present, most personalized recommendation technologies aim at improving the forecast accuracy of recommendation algorithms. However, as the scale of recommendation systems grows larger and larger, the number of customers and items reaches millions and moreover. The overlapping of any two items and choices among customers is so little, it comes to the bottleneck in the forecast accuracy of recommendation systems:the data sparsity which leads to the rapid decline of recommendation quality. This paper studies on this problem and introduces several recommendation algorithms that have improved traditional recommendation technologies. And then based on this, two recommendation methods are proposed:CCFHU, a hybrid of local optimization for recommendation algorithm and CUCF, a collaborative filtering algorithm based on item features and local optimization.The present study includes the following:1. Introducing several traditional recommendation algorithms with their improved ones, focusing on the introduction of the collaborative filtering recommendation algorithm and its advantages and disadvantages.2. On overcoming the data sparsity problem, the method of local optimization option is used to select neighbors as a reference of the target group. This effectively lowers the prediction errors, and makes the errors converge a fixed value to some extent. Moreover, to deal with the inaccurate prediction results particularly existing in collaborative filtering prediction, a content-based method is applied to amend the results and a hybrid of local optimization for recommendation algorithm CCFHU is put forward.3. Proposing a Laplasse smoothing method to optimize the problem that there are errors in the similarity of item features, that will provide a relatively accurate similarity between items when the item features are quite few. Then presenting a collaborative filtering algorithm based on item features and local optimization selection, namely CUCF, on account of the idea of selecting neighbors as a reference in local optimization. From the experimental contrast of CUCF and another four recommendation methods, it turns out that the prediction accuracy of CUCF is raised 7.1% to 15.5%. Thus it proves that the CUCF algorithm can effectively reduce the negative impact of data sparsity, and can achieve better results in terms of forecast accuracy.