Shilling-attack-tolerant Personalized Recommendation Algorithm Based On Matrix Completion

Modern society has entered the era of big data.Huge amounts of data has led to the problem of information overload,which brings about increasingly difficult access to effective information,reduces social efficiency and hinders the development of the society.To solve this problem,the academia has proposed various recommendation algorithms to produce recommendation results through the analysis of users'preference and push it to the users,so as to relieve the information overload problem.Existing recommendation algorithms,however,still face many challenges.In the complex recommended scenario,recommendation accuracy remains a matter of ascension.This paper studied some common problems of the recommendation system and corresponding solution is presented on the base of low rank matrix completion theory.This paper firstly aimed at shilling attack in recommender systems,which is the phenomenon that malicious users use the openness of the recommender system to insert false ratings so as to affect recommender results,and proposed a shilling-attack detecting algorithm based on norm regularized matrix completion,which regards the ratings of attack users in the rating matrix as structural row noise that affects approximate low rank of matrix and models them with_2,1-norm.Experimental results show that the proposed shilling-attack detecting algorithm can detect the shilling-attack users under careful camouflage at a relatively high accuracy and removing the distractions of shilling-attack before recommending can effectively improve the recommendation effect.In addition,this paper fully exploited attribute information as constraint outside of the ratings data to solve sparseness problem of ratings data in recommender systems and further proposed a shilling-attack detecting algorithm based on attribution facilitated matrix completion,which greatly increased the shilling-attack detecting accuracy compared to previous algorithm without attribute information.At the same time,in the aspect of the prediction of ratings,when there is no ratings for new users and items to describe the profiles of new users and items,this article can also model users and items by introducing the attribute characteristics of users and items so as to produce the appropriate prediction of ratings,which provided a solution for the cold start problem in the field of recommender systems.