Comparative predictive modelling of machine learning techniques for network traffic analysis

Network traffic analysis plays a critical role in intelligent network management by providing insights into communication patterns, resource utilization, and network performance. The increasing complexity and volume of modern network traffic necessitate predictive modelling techniques capable of capturing nonlinear relationships among diverse traffic characteristics. This study presents a comparative predictive modelling framework for network traffic analysis by reformulating the prediction task as a supervised regression problem using the benchmark UNSW-NB15 dataset. Unlike its conventional use for intrusion detection, the dataset is transformed into a model of continuous network traffic metric derived from quantitative flow characteristics, enabling the prediction of traffic behaviour rather than attack classification. Five state-of-the-art machine learning regression algorithms, namely Support Vector Regression (SVR), Random Forest Regression (RFR), XGBoost Regression, Light Gradient Boosting Machine (LightGBM) Regression, and Multi-Layer Perceptron (MLP) Regression, are developed and evaluated to investigate their predictive capability for network traffic analysis. The comparative evaluation is performed using multiple regression performance metrics, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Relative Accuracy, Tolerance Accuracy, and the coefficient of determination ( \(\:{R}²\) ). The numerical analysis shows that XGBoost demonstrates a higher performance with \(\:{R}²\) of 0.9969, accuracy of 0.9921, closely followed by LightGBM with \(\:{R}²\) of 0.9957 and accuracy of 0.9906. The ensemble model split its weight almost evenly between these two models (47.5% each), achieving R² of 0.9965 and tercile accuracy of 99.12%, competitive with the best single model, indicating high precision and generalization ability and pointing to the stability and applicability of the model in real-world settings.