Meleshko Y. The methodology of ensuring the stability of recommendation systems to destabilizing factors in computer networks

The thesis is devoted to solution of an actual scientific and practical problem of accuracy increasing for propositions of recommendation systems in the conditions of destabilizing factors in computer networks on the basis of development of models and methods of synthesis of a subsystem of stability ensuring. It is shown that today recommendation systems in computer networks are increasingly used for promotion of content, goods and services. Such systems are tools for automatic generation of recommendations based on study of personal needs of website users. The analysis showed that the vast majority of existing models and methods are vulnerable to internal and external destabilizing factors in computer networks. Ensuring the stability of recommendation systems to the action of destabilizing factors is an important condition for improving their accuracy. Examples of internal destabilizing factors in recommendation systems would be cold start problem, constant cold start problem, filter bubble problem of and problem of insufficient quantity and quality of input data, as well as problem of permanent change of user preferences over time. A method for determining the dynamics of probabilities of recommendation system in its possible states using the mathematical apparatus of Markov and semi-Markov processes is developed. This allows to determine the probabilities of staying specific recommendation system in its possible states at any moment of time. A mathematical model of a stable recommendation system is developed on the basis of the proposed method of determining the dynamics of probabilities of staying the system in its possible states, allowing to optimize the total cost of system maintenance in the conditions of internal destabilizing factors. The method of collaborative filtering is improved. This method differs from the existing ones in using production rules to determine user similarity and using user activity indicators to form recommendations. This allowed to increase system stability in conditions of insufficient input and cold start. A mathematical model of information security subsystem of a stable recommendation system is developed on the basis of the proposed method of determining the dynamics of probabilities of staying the system in its possible states, this allowed to determine the optimal frequency of checking for information attack and bot profiles in the system. A method of software simulation modeling of users and objects of the recommendation system of a social network or web resource based on existing and developed methods of modeling the structure of complex networks and methods of modeling user behavior is developed. This allowed to generate input data for testing the quality of algorithms for forming recomendations. A method of detecting an information attack on the recommendation system based on the analysis of trends of the ratings of objects is developed. This allowed to reduce the cost on monitoring system security by eliminating the need of search for bots in the absence of signs of attack. A method of detecting botnets in the recommendation system based on graph clustering and analysis of user actions for ensuring stability of the system to external destabilizing factors is developed. This allowed to detect botnets and differ them using sets of attack objects. Algorithms for software simulation modeling of users and objects of the recommendation system are developed. These algorithms allowed generating input data for testing algorithms that form recommendation lists. Improved collaborative data filtering algorithms are developed to generate more accurate lists of recommendations for web-resource users based on production rules and the use of user activity metrics. Algorithms for detecting the presence of an information attack on the recommendation system based on the analysis of trends in the ratings of system objects are developed. Algorithms for detecting separate bot profiles based on neural networks and algorithms for detecting bot networks in the recommendation system based on graph clustering and analysis of user activities are developed. A technique for obtaining analytical relations for calculating the probabilities of a stable recommendation system in its possible states at any time to optimize the frequency of recalculation of input data for the formation of recommendations are developed. The assessment of credibility and efficiency of the proposed methods and models for increasing the stability of recommendation systems was carried out. Thus, the results obtained in the dissertation allow to increase stability of recommendation systems to internal and external destabilizing factors and this allows to increase accuracy and other quality indicators of recommendation lists creating.