A Study Of Recommender Systems Based On Collaborative Filtering Algorithm

Nowadays, how to effectively filter the information overload and to find the useful data becomes a hot issue in the direction of web data mining. Faced with the Internet information overload, recommender system is a very effective tool. Recommender systems are usually classified into the following categories: content-based recommendations, collaborative filtering recommendations and hybrid approaches which combine collaborative and content-based methods. Among them, the collaborative filtering method is one of the most successful and most widely used method, which had made rapid development both in theory and in practice. Collaborative filtering recommendations recommend items to people with similar tastes and preferences, based on the users' explicit preferences or implicit preferences. Howere, there are still issues such as data sparsity problem and similarity measurement which seriously affects the quality of collaborative filtering recommendations. In this work, some improved collaborative filtering algorithms were proposed to deal with these issues. The main contributions of this paper are listed as follows :(1)We proposed an improved Jaccard Uniform Operator Distance(IJac UOD) method which is used in collaborative filtering algorithm. In the recommendation system, the key step in the collaborative filtering is the determination of nearest neighbor. In collaborative filtering algorithm, similarity measurement is used to measure the similarity of two users or items. Then according to the similarity we computed, we choose the users who are the most similar to the objective user as the nearest neighbor. Therefore, in order to choose the nearest neighbor efficiently, we proposed an improved Jaccard Uniform Operator Distance(IJac UOD). Compared with the comventional Cosine similarity strategy, Pearson coefficient, the method we proposed properly solves the influence of different length vector difference on similarity.(2)We proposed a mean-based variable weight collaborative filtering algorithm. We use the item average ratings to insert into the user-item rating matrix where is none, in order to increase the density of the user-item rating matrix. After getting a no missing value user-item rating matrix, we use various weighted similarity method to calculate the similarity between users, then look for the objective user's nearest neighbors and predict the rating of items which the objective user didn't rate.(3)We proposed a SVD-based variable weight collaborative filtering algorithm. We use the singular value decomposition(SVD) strategy to predict the missing data in the user-item rating matrix. After getting a no missing value user-item rating matrix, we use various weighted similarity method to calculate the similarity between users, then look for the objective user's nearest neighbors and predict the rating of items which the objective user didn't rate.