Attack Detection Method Based On Time SFM Factors In Recommender System

With the rapid development of e-commerce, the practicability of e-commerce recommender system in handling information overload is gradually accepted and recognized. The key for successful recommender system depends on the external user information, which provides shilling attackers who interfere with normal operations and want delusion recommendation the possibility of theory. And in real e-commerce environment, the commodity producers certainly hope that the recommender system can recommend more products to its users more times, to increase product sales and market share, which provides shilling attackers the possibility of realistic interests. So, improving the attack detection is in favor of ensuring the accuracy, objectivity and impartiality of recommender system.The current attack detection researches on E-commerce recommender system focus on the evaluation value of user's rate items, the number of user's rate items, the changes and differences in the rate number of items and so on. However, during the invasion and the formation of recommendation attacks in the system, attacks must involve the attacker's kinds of information, which is not only limited to user profiles, but also reflected in other characteristics and property on user behavior. And Chirita model has inefficiency on the attack with great number of users and high amount of filler items. It opens a door for the perspective of the analysis and detection of recommendation attacksaccording to it above, this paper put forwards an attack detection method based on time SFM factors representing the characteristics of time gaps in recommender system, which represents the time attributes of user behavior. Firstly by way of getting rid of non-normal time gaps of user evaluations, we generate a relatively wide range of time series and build up a ascending sequence with those of all users. Then, we obtain the reference value for time gaps of the user evaluation, depending on time gaps of evaluations which belong to other users near its upper and lower in the sequence. Finally, according to user's own time gaps of evaluations and his reference value, we constitute the time SFM factors, including Span property from the general angle of group users: the degree of deviation between average value of user's own time gaps of evaluations and his reference value; Frequency Property from the qualitative angle of Single user: the degree of contrast between the number of time gaps and that of time gaps which are greater than his reference value; Mount property from the quantitative angle of Single user: the degree of difference between average value of time gaps of user's evaluations which are more than his reference value and that which are less than his reference value. By analyzing the changes and principles of the time gaps between different rates, it abounds detection elements and extends detection accuracy.In connection with expected goal of the Designed detection model and through simulation experiment of the data set, it shows that the proposed time SFM factors indeed can be put into use for testing the recommendation attack and the attack detection method based on time SFM factors in recommender system can effectively prevent their interferences with TopN recommendation lists for users, and ensure top-quality service of recommender system.