METHOD OF IDENTIFICATION BOT PROFILES BASED ON NEURAL NETWORKS IN RECOMMENDATION SYSTEMS

• Main Authors: Yelyzaveta Meleshko, Oleksandr Drieiev, Hanna Drieieva
• Format: Article
• Language: English
• Published: National Technical University "Kharkiv Polytechnic Institute" 2020-06-01
• Series: Сучасні інформаційні системи
• Subjects: recommendation systems; information attacks; information security; Internet bots; neural networks; data clustering
• Online Access: http://ais.khpi.edu.ua/article/view/204311

The subject matter of the article is the process of increased the information security of recommendation systems. The goal of this work is to develop a method of identification bot profiles in recommendation systems. In this work, the basic models of information attacks by the profile-injection method on recommendation systems were researched, the method of identification bot profiles in recommendation systems using the multilayer feedforward neural network was developed and the experiments to test the quality of its work were conducted. The developed method is to identify bot profiles that attempt to change item ratings in a recommendation system in order to increase the occurrence frequency of target items in recommendation lists to all authentic users, or to certain segments of authentic users. When removing bot profiles' data from the database of the recommendation system before generating recommendation lists, the accuracy of the system and the correctness of recommendations are significantly increased, and authentic users get protection from information attacks. Random, Average and Popular attacks were used to model the attacks on a recommendation system. To identify bots, their ratings for system items were analyzed. The experiments have shown that the neural network that analyzes only the numbers of different ratings in a profile, detects bot profiles with high accuracy, that use Random attack regardless of the number of target items for each bot. At the same time, the developed neural network can detect bots that use Average or Popular attacks only when they have several target items. Also, the results of the experiments show that type I errors, when the system identifies authentic users as bots, is very rarely appear in the developed method. To improve the accuracy of the neural network, there can add to analysis also other data of user profiles, such as the timestamp of each rating and as segments of items, which was rated.