Research On Hybrid Collaborative Filtering Algorithm Based On Multi-Features Fusion

With the rapid development of Internet, information overload problem has become increasingly prominent: the users cannot find needed information quickly and efficiently from a large number of network resources. Recommendation system helps solve this problem by recommending items to users based on their previous preferences and now it has applied widely in many domains.Collaborative filtering is one of the most successful techniques in recommendation system because it is easy to implement and work well in many real-world situations. Collaborative filtering predicts the active user preference for goods or services by collecting a historical data set of users’ ratings for item, the underlying assumption is that the active user will prefer those items which the similar users prefer. However, the data is quit sparse, which makes the computation of similarity and Similar Neighbors Selection imprecise, and consequently reduces the accuracy of recommendations. Besides, with the gradual increase of customers and products, the problems of real-time and scalability also limit the application of collaborative filtering techniques. For these problems, the main contributions of this dissertation include the following aspects:(1) To address the problem that traditional similarity methods compute inaccurately on sparse data, firstly we propose an enhanced Pearson similarity method (EPCC)that the common ratings and the all ratings are both taken into account, and the method proved to be more accurate. Besides, in order to alleviate the sparsity of the original rating matrix, a rating-feature fusion Similarity algorithm has been proposed, it uses SVD technology to mine the potential characteristics of the ratings and then the original ratings and the mining characteristics are combined to compute the similarity, especially, the way of middle-fusion in the proposed algorithm can dynamically balance the importance between the ratings and the characteristics. The experimental results show that the proposed Similarity algorithms improve the prediction accuracy effectively.(2) According to the problems existing in the traditional neighbor selection methods, this paper studies on the relationship between the number of the neighbors with the similarity of the neighbors and proposes an improved k nearest neighbors algorithm based on confidence interval. The experimental results show that the proposed algorithm obviously improves the prediction precision with almost the same time complexity.(3) In order to make full use of the users and the items to increase the accuracy of recommendation, a hybrid collaborative filtering algorithm based on the user-item two-dimensional neighbor selection is proposed. The proposed hybrid algorithm combines the user and the item in the section of similar neighbor selection, and the selected neighbors’number would be a part of weighing for hybrid prediction. Besides, for reducing the difference between users’scale ratings and the real ratings, and thereby increasing the prediction accuracy, we present a probabilistic interval matching prediction method, it firstly predicts a missing data probability rating interval and then the final missing data rating is produced by matching the interval. The experimental results show that our proposed hybrid algorithm outperforms other hybrid algorithms and it greatly increases the recommendation accuracy in the situation of data sparsity.