Study On Collaborative Filtering And Privacy Protection Mechanism In Recommender Systems

Recommender Systems can help users to discovery information of their interest, and solve the information overload problem efficiently. They first analyze histories of users’ behaviors and achieve appropriate models for prediction, then produce recommendations for users. Researches on recommender systems mainly focus on the rating prediction problem, where an important way to solve this problem is collaborative filtering. How to improve the prediction accuracy is the key issue in the literature. Besides, in recent years, the privacy protection problem in recommender systems gradually becomes another key issue in collaborative filtering. It puts the growth of recommender systems at risk, because recommender systems may modify users’ratings to mislead others as interest dictates. Therefore, a privacy protection mechanism is needed to ensure accurate and effective recommendation; and preotect users’ratings against the misuse of recommender systems.This paper mainly deals with two-fold issues:one is to improve the prediction accuracy of collaborative filtering algorithms. The other is to protect users’ratings from misusing by recommender systems. Firstly, two types of effective collaborative filtering algorithms which are used in Netflix prize were studied:the neighbourhood-based model (KNN) and the matrix factorization model (MF). Then an approach of combining the result of different models and model combination strategy were investigated by analyzing two public released data sets-MovieLens and WSQ data sets. Finally, a privacy protection mechanism is introduced. It applies homomorphic encryption to encrypt users’ratings, and processes them using KNN to generate recommendations. Our contributions are as follows:1. We have implemented some classic models of collaborative filtering algorithm, including iKNN、Funk-SVD、Biased SVD and SVD++. Then we combine them by the way of simple linear regression. Finally, the analysis of model combination strategy is described through two kinds of experiments. The experimental results show that the accuracy improves as more models are combined, and the combination of different types of models gets more accurate than the one of the single types of models. Therefore, excellent systems can be built with just a few well-selected models.2. We have researched a privacy protection mechanism based on Paillier and DGK homomorphic cryptosystems. By interacting with a semi-trusted third party, recommender systems process encrypted ratings and find similar neighbors to generate recommendations. Compared with the similar users selection strategy based on threshold, this paper proposed a similar users selection strategy based on interval which can guarantee for each user to find a number of similar users.