Predicting Network Performance Degradation in Wireless and Ethernet Connections Using Gradient Boosting, Logistic Regression, and Multi-Layer Perceptron Models

Abstract

This study explores predicting network performance degradation in wireless and Ethernet connections using three machine learning algorithms: XGBoost, Logistic Regression, and Multi-Layer Perceptron (MLP). Key metrics, including accuracy, precision, recall, F1-score, and AUC-ROC, were employed to evaluate model performance. The MLP classifier achieved the highest accuracy (98.7%) and AUC-ROC (0.9998), with a precision of 1.0000 and recall of 0.8622, resulting in an F1-score of 0.9260. Logistic Regression provided reasonable baseline performance, with an accuracy of 93.67%, AUC-ROC of 0.9565, and an F1-score of 0.5992, but struggled with non-linear dependencies. XGBoost showed limited utility in detecting degradation events, achieving an F1-score of 0 despite a perfect AUC-ROC (1.0), indicating sensitivity to imbalanced data. Through hyperparameter tuning, MLP demonstrated robustness in capturing complex patterns in network latency metrics (local_avg and remote_avg), with remote_avg emerging as the most predictive feature for identifying degradation across both network types. Visualizations of latency dynamics demonstrate the higher predictive relevance of remote latency (remote_avg) in both network types, where spikes in this metric are closely associated with degradation. The findings underscore the effectiveness of using latency metrics and machine learning to anticipate network issues, suggesting that MLP is particularly well-suited for real-time, predictive network monitoring. Integrating such models could enhance network reliability by enabling proactive intervention, crucial for sectors reliant on continuous connectivity. Future work could expand on feature sets, explore adaptive thresholding, and implement these predictive models in live network environments for real-time monitoring and automated response.