| With the rapid development of network and information techniques, web applications can offer more and more information and services than ever before. As a result, users must to face the mass data with lots of useless informations on the web, so-called"information over-load problem". Facing the huge amount web resources, recommendation system is a potential personalized technique for solving the above issues, which can adjust the content and means of services by tracking users'interests.Collaborative filtering is one of the most widely used and successful methods for recommendation, which has been made fast development in theoretical research and applications. Collaborative filtering computes the similarity of users via their personal profiles, generating recommendations for the current user by the preference information of his neighborhood. However, collaborative filtering has got challenges, such as data sparsity, difficulty of similarity measurement, low recommendation precision and scalability issues with the fast growth in the amount of users and items. In this work, some improved collaborative filtering algorithms were proposed to deal with these issues. The main contributions of this dissertation are as follows:1) Most existing calculations of similarities suffer from data sparsity and poor prediction quality problems. For this issue, we propose a similarity measurement algorithm based on entropy. The entropy is computed by the difference of two users'ratings, and we also consider the size of their common rated items, the size is bigger, the weight of their similarity is higher. Experiments show that the algorithm effectively solves the problem of the inaccuracy of similarities in data sparsity or small size neighborhood environments, and outperforms other state-of-the-art CF algorithms and it is more robust against data sparsity.2) Traditional collaborative filtering methods just use user-item rating matrix to generate recommendations, and lead to difficult to computer the similarity because of the data sparsity. We propose a hybrid collaborative filtering algorithm combining the rating matrix and item attributes. First, we design a user similarity measurement method by computing the user's preference to different item attributes, this approach is consistent with the true relationship between users, and also can effectively alleviate the issue of rating matrix sparse. Then, when computing the similarity of two users, we combine the Pearson correlation and the items attribute preference similarity, with a weighting coefficient"w"to balance the importance of two parts. Experiments show that this algorithm effectively solves the problem of data sparsity, and outperforms better when the sparsity is more serious, compared to the traditional CF algorithms.3) Memory-based CF algorithms have the weakness of low real-time ability and scalability. For these issues, a SVD-based K-means clustering CF algorithm is proposed. Traditional clustering-based CF algorithms have low recommendation precision because of data sparsity. So we first fill the missing ratings by SVD prediction, and then implement k-means clustering in the filled matix. This algorithm overcomse the data sparsity issue via SVD and keep the advantage of clustering, such as good real-time ability and scalability. Experiments results show that this algorithm outperforms Pearson CF, svd CF and k-means CF.4) For the issue of single CF algorithm performing low recommendation precision, we propose an adaptive AdaBoost.RT ensemble learning algorithm. First, the base regression predictor is formed by minimizing the error function of user's predicting ratings via gradient descent algorithm. Then, we introduce an adaptive error parameter, which has statistical property and can be adjusted automatically by the predict error, instead of original parameter. Experiments results demonstrate that this ensemble learning algorithm can rise the performance of single CF model significantly.
|
PDF Full Text Request