Robust Non-negative Matrix Factorization Recommendation Algorithm Based On Shilling Attack Detection

Due to the characteristics of collaborative filtering technology and the openness of the Internet,the collaborative filtering recommendation system is relatively fragile and can't resist the impact of the shilling attack.Although the existing research has proposed some robust recommendation algorithms,but these algorithms either the resistance for shilling attack is not strong enough,the performance of robustness is poor;or to improve the robustness of the algorithm at the expense of the prediction accuracy.Aiming at the above situation of robust recommendation algorithm,we propose a robust non-negative matrix factorization algorithm based on shilling attack detection to improve the robustness of the system.Firstly,we improved the traditional singular value decomposition variable selection algorithm and proposed a shilling attack detection algorithm based on singular value decomposition and weighted deviation from mean agreement detection attribute.First,we improved the strategy of the traditional singular value decomposition variable selection algorithm to determine the number of suspect users,reduce the value of the flag,and improve the accuracy rate in the case of guaranteed recall rate.Secondly,we proposed a robust non-negative matrix factorization algorithm based on shilling attack detection by combining the improved singular value decomposition variable selection algorithm and the weighted deviation from mean agreement detection attribute.The algorithm uses the VSSW algorithm to detect the rating data before iterating the update,and obtain the suspect user set.And then filter out the users who are included in the suspect user set in the process of non-negative matrix factorization,so that it does not participate in the iterative process,thus offsetting the malicious effect of the attack profile on the recommended results and improving the robustness of the algorithm.Finally,for the algorithm mentioned in this paper,the simulation experiment is carried out on the MovieLens 1m dataset,and the experimental results are compared with some existing algorithms.Compared with the existing algorithms,the algorithm proposed in this paper can ensure that the robustness of the algorithm is improved without loss of recommendation accuracy.