Abstract
The present work explores the development and application of the method of indirect monitoring of telecommunication network quality based on the analysis of Point-to-Point Protocol over Ethernet (PPPoE) session parameters using machine learning methods as a key indicator of the network failures, the use of the K coefficient is justified based on the dynamics of PPPoE Active Discovery Termination (PADT) packets and the number of active PPPoE sessions. The paper describes the stages of data collection and preprocessing, including the conversion of session indicators from a “wide” format to a “long” format for ease of analysis. A statistical analysis of the significance of attributes (Analysis of Variance (ANOVA)-test, correlation analysis) was carried out, based on which a limited set of informative parameters of PPPoE-sessions (e.g., connection duration, frequency of disconnections, volume of transmitted data, connection establishment time) was selected. Linear Regression, Ridge, Lasso, Random Forest, and Support Vector Regression (SVR) models were trained and comparatively evaluated on these attributes to predict the K value. The symbolic regression experiment provided an analytical formula to confirm the correctness of the selected K value. The comparative analysis by the Mean Squared Error (MSE) and Coefficient of Determination (R²) metrics showed the advantage of Random Forest model (R2 ≈ 0.90, MSE ≈ 0.0001), which indicates the high efficiency of the proposed approach. The significance of the study lies in demonstrating the possibility of the early detection of the network quality anomalies without a direct analysis of the traffic content, which increases the efficiency of monitoring the quality of telecommunication services.