Research On Probability Output Method For Integrated Recommendation Attack Detection Modeel

With the rapid development of electronic commerce, the problem of "information overload" becoming more and more serious, in order to solve this problem recommendation system appeared. But the recommendation system depends on external user information,which is very sensitive to user information. It is susceptible to fake user profiles. Due to the user profiles attacks, the accuracy of the results of recommendation system is very low. The existing recommendation attack detected methods causing large number of real users were mistaken, and deleted from recommendation system. Therefore, how to eliminate the system to filtered the real user profiles has been crunch time. We analysis the methods of attack detection at home and abroad, then studied collaborative filtering recommendation system for attack detection algorithms.Firstly, aiming at the problem of user information filtered in attack detection algorithms on recommendation systems. Through analysis the difference between real user profiles and false user profiles. We builded the probability distribution of the user profiles based on Gauss Mixture Model. Then we proposes a probability output method for based classifiers, let attack detection methods has the probability output ability.Secondly, in order to ensure accuracy of the probability output for the integrated recommendation attack detection. This paper we analysis evidences conflict, which is generated by probability vectors. Then put forward a method to solve the evidences confliction, which based on packet optimization algorithm. Through this algorithm the evidences can use the D-S theory directly.Finally, we give the experimental evaluations and analysis of the probability output method we proposed in this paper, and compare the performance between we proposed methods and other existing methods, then make the conclusions and prospects the further search.