The Research On Robust Recommendation Model Based On User Rating Matrix Block

With the rapid development of Internet, recommender system has been playing amore and more important role. However, due to the openness of recommender system andthe purpose of business competition, some malicious users inject a large number of falseuser profiles into the system deliberately to affect the recommendation quality. Theexisting recommendation models based on matrix factorization expose low robustnesswhen facing the shilling attacks, and the robustness is improved at a cost of accuracy.Moreover, with the increase of users and items constantly, the existing recommendationmodels need to be updated by reconstructing, which is always time consuming. Aiming atthe above problems, on the basis of the present research at home and abroad, this paperhas a further research on how to improve the robustness of the recommender system onthe guarantee of recommendation accuracy.Firstly, the most widely-used collaborative recommendation models are vulnerable toattack profiles and the robustness is low. To this end, we propose a robust recommendationmodel based on user rating matrix block and modified LTS-estimator. We construct userrating matrix blocks by using user rating matrix block algorithm based on k-medianclustering. We apply the modified LTS-estimator to matrix factorization model in order toproduce user feature matrix and item feature matrix. Based on above, we construct arecommendation model to generate recommendations for the target users.Then, the existing recommendation models improve the robustness at a cost ofrecommendation accuracy and are always time consuming to update. Aiming at thatproblem, we present a robust recommendation model based on incremental clustering andmatrix factorization. We use user rating matrix block algorithm based on k-medianclustering to construct the user rating matrix blocks. Also, we use matrix factorization foreach block, based which to construct the recommendation model. For the new users oritems, we use incremental clustering algorithm to update the user rating matrix blocks. Weadopt the linear regression approach based on weighted entropy to realize the localparameter estimation. Finally, we organize experiment between the proposed robust recommendationmodels and other related collaborative recommendation models to compare theexperimental performance and analysis.