Research And Realization Of Personalized Recommendation

Personalized recommendation system becomes popular in the context of information explosion. It consists of collaborative filtering and content-based filtering.Information is recdmmended to the target user based on recommendations of other users who has similar interests with him in collaborative filtering. The magnitudes of items and users in the system results in the extreme sparsity of user rating data, which makes it difficult to find neighbors effectively. In this paper, a recommendation algorithm-Matrix Partition and Interest Variance Based Collaborative Filtering Algorithm is proposed. It partitions the huge matrix into some submatrixes in order to reduce the scale of searching nearest neighbors. While partitioning the matrix, it first classfies resources based on explicit classfied system of resources, and then clusters those classes which contains more resources relatively because they have no explicit classfied system. Moreover, the concept of interest variance is adopted to improve the veracity of searching nearest neighbors. It proves that this method can obtain a better predictive precision, compared with traditional recommendation algorithm and algorithm which don't take into account interest variance.Information is recommended to the user based on the similarity between information and his own interests in content-based filtering. In traditional recommendation system, user profile is denoted with a set of keywords which denotes the interests of users and they are isolated. Owing to not considering the latent relation among words, the quality of the system is weakened. In this paper a new information recommendation system-Carlnfo is introduced. It takes the relations among words into account. It can make scattered keywords into a meshy structure so as to provide an abstract view of users' characters. It can distinguish explicit interests with latent interests. The modification of user profile is made based on user's actual choose. Structuring of ontology is realized semi-automatically. The feasibility of the system is proved by our primary experiments.