Research On Feature Extraction And Detection Appoaches For Recommendation Attacks In Collaborative Filtering

Age of Network economy is coming with the rapid development of electronic commerce. But meanwhile it brings the problem of "information overload". By then, the recommendation technology emerged as the times require. Because of the openness of recommendation system and the sensitivity of user’s information, the recommendation technology risks of information safety. The recommendation quality seriously reduces because some of the malicious users artificially injected a lot of false profile to the system to generate beneficial recommendation to themselves. Therefore, it is necessary for the recommend system to enhance its security system. Based on the comprehensive analysis on the domestic and foreign research, this paper makes a deep research on attack detection algorithms in the collaborative filtering recommendation system.First of all, the general feature extraction methods is not working well on the unknown recommendations, this paper gives two methods to extract the characteristics of the attacks by the introduction of wavelet transform. One is considered from the user’s score distribution, extract the special features of the known type attacks by using all levels of signal energy distribution after the decomposition and reconstruction of the wavelet. The other reduces the dimension of the data using a wavelet decomposition, and then extract the general features of unknown type attacks through the theory of entropy.Secondly, the existing attack detection algorithm can’t detect unknown attacks effectively, is proposed based on the above characteristic value, this paper puts forward a detection algorithm of unknown type attacks based on clustering. Moderate real users from the score database chosen by the random sampling technique with the attack user profile generated by the attack model form the test set. Firstly,we use the general methods of feature extraction to map the test set into the feature space. Users in the feature space clustered by clustering algorithm and small clusters are identified as the attack user profile, detection complete. As the mean attack generating attack profiles, the filer randomly selected in all projects but the target items and the selected items, we presents a known type of attack detection algorithm based on clustering. The algorithm is similar with the above except the special feature extraction method for feature extraction.Finally, user profile Injection Attacks Detection Algorithm and the existing detection algorithms are compared and analyzed in this paper.