Shilling Attacks Detection Methods Based On Rating Sequences

The popularity of the Internet provides people with more convenient access to information,but the Internet is flooded with information,people need to spend a lot of time to obtain the information they need,such as information overload phenomenon plaguing people selection of effective information.Recommendation system can effectively alleviate the problem of information overload,and the collaborative filtering recommendation system based on collaborative filtering technology and recommendation technology has been widely used in the field of electronic commerce.Due to the openness of collaborative filtering recommender system,malicious users can easily inject false data into the system,which results in the deviation of recommendation results.Therefore,in order to ensure the reliability of the recommendation system,it is necessary to accurately detect the false data in the recommendation system.In this paper,we study the attack detection methods in the recommendation system based on the user's rating habits.First of all,aiming at the problem of low accuracy of the existing attack detection methods,this paper proposes a clustering attacks detection method based on novelty item rating sequence.The method first construct novelty item rating sequence for each user,according to the difference between the genuine profile and the attack profile on the novelty item rating sequence to generate rating number percentage sequence for each user profile,and finally combining with FarthestFirst clustering algorithm to detect the attack profiles.Secondly,in order to improve the detection accuracy of the standard attack and the AoP attack,this paper proposes a detection method based on the classification feature of the rating sequences.The method first construct popular item rating sequence and novelty item rating sequence for each user,and then we propose eight classification features according to the difference of the distribution of the rating items in the rating sequences between genuine profiles and attack profiles,and finally combining with MultiBoost ensemble learning method and ADTree decision tree classification algorithm to detect attack profiles.Finally,the experiments on the MovieLens 1M data set are carried out,and the two kinds of attack detection methods proposed in this paper are validated and compared with the existing several kinds of attack detection methods.