MATHEMATICAL MODELING OF COMPUTER SYSTEM USERS' BEHAVIOR

The large-scale use of computer technology in almost all spheres of human activity has brought more and more attention to the user himself. Knowledge of what actions he performs (or should perform) can be applied in different areas, for example, in security systems, to create a personalized environment for users and so on. Therefore, the task of building models of user behavior of computer systems is relevant. In this paper, we propose a comprehensive user model consisting of interactive and session parts, which in both models to detect deviations from the usual take into account, respectively, dynamic and statistical properties of user behavior for expected user behavior neural networks are used. Thus, the interactive model is based on predicting user commands based on previous commands. Since the choice of neural network architecture is a non-trivial task, it is important to know how much its current behavior depends on the pre-history. In the case of the session-based model, there is a problem with the sample size that is used to train the neural network. The fact is that when the training set size is small, the neural network tends to memorize images locally, which is undesirable. In the session model, the input to the neural network is the data collected during the session as a whole. Accordingly, the size of the training set is directly determined by the number of sessions during which the user's activity was monitored. However, even over a long period of time, this data will not be sufficient for qualitative training of the neural network. Therefore, it is very important to provide a more representative sample of data in the session model for qualitative training of the neural network.