Research On Semi-supervised Detection Method For Recommendation Attack In Collaborative Filtering

With the advent of the era of big data,information overload has attracted more and more attention.Collaborative filtering recommender systems is considered as one of the effective ways to solve the problem of information overload and is widely used in many fields such as e-commerce recommendation.However,the collaborative filtering recommender systems is vulnerable to recommender attacks due to its openness.In this attack,malicious users artificially inject a large number of fake user profiles into the system for their own commercial competition and other purposes,so as to achieve the purpose of changing recommender results.In order to detect the recommender attacks,the researchers proposed three types of detection methods: unsupervised,supervised and semi-supervised,among which the semi-supervised detection methods has the advantage that it can improve the detection performance by using the profiles of a large number of unlabeled user profiles existing in the recommender systems.Based on the comprehensive analysis of the research status at home and abroad,this paper studies and discusses the recommender attack semi-supervised detection methods to further improve the detection performance of semi-supervised detection methods.Firstly,aiming at the low accuracy of existing semi-supervised detection methods,this paper proposes a recommender attack detection method RAD-SFDA based on semisupervised Fisher discriminant analysis to improve the accuracy of semi-supervised detection methods.First,the projection vector is determined by using Fisher discriminant analysis technique combined with the labeled user profiles.Then,principal component analysis is used to extract the global structure from the data set established by the labeled user profiles and unlabeled user profiles.Finally,the best projection vector is determined by combining the above discriminant structure determined by the labeled user profiles and the global structure determined by the all user profiles,and the bayesian classifier is trained in the final projection space,and the trained bayesian classifier is used to detect the data in the test set.Secondly,although the proposed method RAD-SFDA improves the accuracy of the semi-supervised detection method to a certain extent,the detection stability of the method is not high,mainly manifested in the method's poor ability to identify the recommended attacks with high filling sizes and high attack sizes.Aiming at this problem,this paper proposes a semi-supervised detection method based on ensemble study SemiBoost RAD-SemiBoost,this method first according to the similarity between the user profiles for unlabeled user profiles giving pseudo label and sampling operation,then the sample unlabeled user profiles with pseudo label the user profile and labeled user profile training individual classifier together,finally,the use of integrated learning thought,individual classifier combination of multiple iterations into the final ensemble classifier,using the final classifier to detect recommender attacks.Finally,on the Movie Lens 10M data set of the collaborative filtering recommender domain standard,an experimental comparison is made between the proposed method and related work to verify the effectiveness of the proposed method.