Study On Collaborative Filtering Recommendation Algorithm Based On Feature Vectors

recommendation system has attracted wide attention and has been used widely. It is also the effective way for customers to the needed information from the increasing tremendous information. By the use of e-commerce personalized recommendation system, e-commerce websites offer personalized service and simulate the behavior of shopkeeper to help customer finish the process of buying.The e-commerce personalized recommendation system analyses the correlation between users and information, make use of the similarity to find the users'interesting in mining of the plenty of information sets. The essence of recommendation system is filtering. There have been several kinds of recommendation algorithm proposed and adopted, which include population statistic data based, content-based, collaborative filtering, and hybrid method. Among these technologies, collaborative filtering algorithm is regarded as the most successful and the most widely used recommendation algorithm. With the increasing e-commerce scale, collaborative filtering algorithm faces the some challenge, such as data sparseness, scalability, cold-start etc. This thesis made a lot research on collaborative filtering algorithm. The main contributions of this thesis are summarized as follows:Firstly, Observing of the law of users' rating. The advantage and deficiency of existing consistency rating methods are evaluated and analyzed. In order to solve the inconsistency of users' rating criteria, an improved method are proposed. Secondly, the conventional calculation of similarity focuses on the calculation of whole rating matrix. Matrix sparseness affect the accuracy of the calculation, and the calculation also do not take the feature of goods into account. Therefore, the conventional calculation of similarity exist poor recommendation quality and bad real time feature. This thesis proposes a novelty recommendation algorithm based on feature vector. The algorithm scans the rating matrix and the feature of goods by using statistical analysis techniques and gets the profiles of users and goods, which the feature vector is stored in, then calculate the similarity. The algorithm also reduces the computing complexity, improve the data sparseness and raise the recommendation quality.Thirdly, analyzing and improving the existing recommended processes. The proposed algorithm makes the optimization of some of the steps and reduces the unnecessary computation, and improve real-time by using the optimization strategy with less impact on recommendation accuracy. Algorithm can be divided into three stages: initialization, offline update, on-line. In the initialization stage, build the profiles of users and goods and calculate the similarity between them; in the offline update stage, update the similarity between users and the nearest neighbor. In online stage, make a recommendation for users without scanning all goods by using proposed improved algorithm. And increase the updated profiles when users rate new goods.Finally, Based on the data sets collected by Minnesota university, which are widely used in current researches, this paper design some contrast experiments, analyze and compare the recommendation accuracy of the Cosine, Pearson, Off-Cosine, demonstrate that improved algorithm is more accurate and real-time than the traditional collaborative filtering recommendation algorithm. The experimental results indicate the algorithm get the desired recommendation accuracy and real-time.