Using Outlier Detection To Improve Collaborative Filtering Recommendation Algorithm

The rapid development of e-commerce makes the number of network consumer continue to increase and the network information is greatly springing up. Merchandise recommendation technology can provide users with personalized service according to the different requirements, so it has been widely used in electronic commerce system. Collaborative filtering recommendation algorithm is one of the most effective algorithms currently used, but the algorithm still has some problems, such as being easily influenced by noisy data, and data sparsity caused by the increasing number of users and items in the system. Based on the traditional collaborative filtering recommendation algorithm, this paper puts forward some improving methods. The main works of this paper are as follows:(1) Proposing an improved method to calculate user rating similarity. The calculation of user’s rating similarity is the most critical step of the recommendation system, and it determines the neighborhood selection of the target user. Similarity is calculated by user’s ratings and the item mean. Compared with the traditional method, this method can measure similarity between user’s ratings well.(2) Using density-based local outlier detection algorithm to remove the noisy data from the target user’s neighborhood. After determining the neighborhood of target user, outlier factor for each user is calculated by using density-based local outlier algorithm, users whose outlier factor is bigger than the preset threshold are removed, and then the rest of users are recommended to target user.(3) Alleviating data sparsity by filling the rating method based on user characteristics. If users’ history rates were too sparse, it is difficult to provide accurate recommendation for users. Before generating recommendation by collaborative filtering algorithm, the items which have no rate in the rating matrix is filled based on the principle that users with the same characteristics may have similar ratings, then data sparsity is alleviated and the forecast rate is more accurate.(4) By doing some experiments on the Movielens dataset, the experiment results show that the improved algorithm can fatherly improve the quality of recommendation system.