Personalized Recommendation Technology Research Oriented Sparse Data

With the development of Internet and e-commerce, there is great increase in Internet information. Because of the "information overload" phenomenon, it is difficult for users to find the information they need. Even though search engine can save searching time and locate the information to meet the needs of users quickly, it lacks the ability of customizing personalized search results. Personalized recommendation technology provides personalized recommendation services to meet the needs of users and solves the problem of information overload effectively according to users’ behavior history information. As the scale of personalized recommendation system is larger and larger, the relative user behavior history information is less and less, and the rating matrix is extremely sparse, which leads to a reduction of the quality of the recommendation.In order to alleviate the impact of data sparse, the dimension reduction is considered. It uses feature extraction or machine learning methods to reduce the dimension and also use the principles of data compression to reduce the dimension of the system data. Existing recommendation technology discards some important information in the process of reducing dimension, which makes the accuracy of similarity calculation and neighbor searching is not high, thus affecting the recommended effect. In the view of problems of existing recommendation technology on sparse data, the main work of this paper is as follows:(1) When the rating matrix is extremely sparse, the user similarity calculated through existing similarity measurement method is inaccurate, which leads an inaccuracy when finding the nearest neighbors based on the users’ similarity matrix and calculating the predicted value of rating. Therefore, we proposed a personalized recommendation algorithm based on clustering and trust(PRACT). It defined a trust similarity, then combined users’ trust based on scores mode with traditional similarity measure algorithm based on user ratings. We took experiments on MovieLens datasets, and the experimental results showed that the user similarity calculated by the proposed algorithm was more accurate, which could help to find the nearest neighbor users appropriately and fill the predicted values of rating more accurately. In the case of avoiding users to rate maliciously, it improved the accuracy of the user similarity calculation in the case of sparse data and the recommendation accuracy.(2) Currently the ratings data of recommended system is high dimensional and sparse, such as multimedia data, shopping data and so on. Most existing recommendation algorithms for high-dimensional sparse data just focus on the dimensional reduction, but they discard too much useful information in the process of reducing dimension, which results in a decline in the precision to find neighbors. In order to solve this problem, we proposed a personalized recommendation algorithm based on sparse subspace clustering(PRASSC). It used sparse subspace clustering algorithm to cluster users, reduced the discard useful information, then found relevant cluster, calculated users’ similarity, found nearest neighbor sets through using K-Nearest Neighbor(KNN) algorithm, and calculated the predicted value of ratings based on the ratings of neighbor sets. We took experiments on MovieLens datasets, and the experimental results showed that the proposed algorithm could actively ease the problem that finding neighbors inaccurately under the condition of sparse data and improve the recommendation quality.