Research On Robust Recommendation Algorithm Based On User Reputation

The rapid development of Internet technology has promoted the rise and prosperity of the e-commerce platform,the amount of information increases exponentially all the time.It is hard for people to find the useful content in the ocean of information.As an effective method to solve the problem of information overload,recommender system plays an important role in filtering information.Collaborative filtering recommendation algorithm is one of the most widely used and successful recommendation algorithms.The performance of collaborative recommender systems relies on the user's rating information.Consequently,recommender systems are very vulnerable to attack by malicious users.Malicious users try to manipulate the recommendation result by injecting the false user profiles into recommender system in order to gain improper benefits.These deliberate behaviors seriously harm the security of the recommender system.To aim at this problem,many scholars have proposed many robust collaborative recommendation algorithms based on matrix factorization model.These algorithms have such problems as low recommendation precision,poor robustness,unsuitable for large-scale attack and high time complexity.This dissertation have made an in-depth investigation and analysis of the domestic and foreign existing robust recommendation algorithms,and mainly researched the problems such as the recommendation accuracy,robustness,time complexity,shilling attack model,shilling attack detection and mining user potential features from the social network in the robust recommendation algorithm based on matrix factorization model.This dissertation includes:1.First of all,we simply described the theory of recommender system and shillingattack,and review the current states and existing problems of recommender system security algorithms both at home and abroad.Then,the robust collaborative recommendation algorithms based on matrix factorization model is introduced in detail.At last,in order to further improve the accuracy and timeliness of the robust collaborative recommendation algorithms,we proposed two effective robust collaborative recommendation algorithms,which is combined the user reputation system with the recommendation system.These two algorithms are suitable for random attack and average attack environment respectively.2.The existing algorithms,which are based on the latent factor model,have poor robustness and low recommendation accuracy under shilling attacks.Thus,we grouped users according to their rating histories and iteratively compute each user's reputation,integrated the reputation information of the use rating group into the matrix factorization model using the current common combination method of reputation and recommendation model,and proposed a robust recommendation algorithm using an iterative group-based reputation(IGRS).It does not destroy the user distribution of the original dataset,can resist larger-scale attacks.Experiments on the real datasets MovieLens 100K and MovieLens 1M show that IGRS algorithm outperforms previous robust algorithms in terms of both accuracy and robustness in the random attack environment.3.At present,the researches on the robustness of the recommender system are mainly focused on improving the ability of against attacks,while the researches on timeliness of robust recommendation algorithms are very few.Usually recommendation algorithms with strong robustness have high time complexity.However,those algorithms with low time complexity are often with poor robustness.To solve this problem,we propose a fast and robust collaborative recommender algorithm based on user's reputation(LSVD)by calculating the user's reputation based on user historical ratings,building the phase function through analyzing the distribution of user's reputation,and combining the user reputation information with matrix factorization model.Experiments on the MovieLens 1M dataset under the average attack environment show that LSVD algorithm can achieve good results on two indexes of execution time and robustness prediction shift,under the premise of high recommender accuracy.