Privacy-Preserving Collaborative Filtering Using Variable Weight Randomized Perturbation Technique

The Internet is closely related to everyday life of each person. Everyone constantly produces large amounts of data every day. As the data accumulated unceasingly, the size of the amount of the Internet data is becoming increasingly unpredictable, which leads to information overload. It is particularly difficult for people to find their needed information from the vast amounts of information. To solve this problem, people come up with two effective methods:search technology and recommendation technology. Both technologies complement each other to solve the information overload problem.Collaborative filtering algorithm as a relatively successful recommendation technology is not very perfect. There are all kinds of problems in Collaborative filtering algorithm, such as data sparse problem, the accuracy of recommendation loss and so on. Meanwhile, with the development of the Internet, the user pays more and more attention to personal privacy. If users want to enjoy the benefits of information system, they have to submit their real data which lead the privacy of their expose. Therefore, under the premise of protecting users’ privacy, to provide the recommendation for users is very important.In order to respond these problems, this paper is divided into the following areas:First, in view of the traditional collaborative filtering recommendation algorithm for the data sparse problem and interest changing with time problem, we proposed optimized collaborative filtering using variable weight technique. If the data base is too sparse, the error of traditional similarity calculation is Serious, what is more, the similarity cannot be calculated. The traditional collaborative filtering algorithm does not consider the characteristics of user interest changing at the same time, which affect the efficiency of recommendation. Therefore, this paper introduces a time drift weight, a joint score rate weight and a balancing function. Experiment results show that the optimized collaborative filtering using variable weight technique can effectively solve these problems.Secondly, in view of the low efficiency of the memory-based collaborative filtering algorithm, we propose clustering model based on the user’s preference. Facts have proven that not all users will contribute to recommend for the active user. So we introduce the modified k-means clustering algorithm to improve the efficiency of recommendation. Experiment results show that clustering model based on the users’ hobbies can effectively improve the efficiency of memory-based collaborative filtering algorithm.Third, in view of the traditional privacy-preserving collaborative filtering using randomized perturbation techniques having a low privacy-preserving and accuracy of recommendation, we propose a privacy-preserving collaborative filtering using variable weight randomized perturbation technique. The noise of traditional technology obeys a particular distribution, and all information for all users are taken to the same perturbation strength, which not only reduces the recommended efficiency, but also reduce the privacy. To solve this problem, this paper paying attention to the privacy attenuation and diversity characteristics, introduce the variable privacy weight. Experiment results show that the privacy-preserving collaborative filtering using variable weight randomized perturbation technique not only improves the users’ privacy, but also improves the accuracy of recommendation. It can make a good balance between accuracy and privacy.In conclusion, the proposed method in this paper effectively alleviate the low efficiency, poor accuracy and weak privacy protection problem of the traditional collaborative filtering recommendation technology.