Adaptive Online Learning Detects Mobile Network Failures.

A new two-stage online learning framework has been developed to enhance the detection of service-affecting failures within mobile core networks. Traditional monitoring methods struggle with the complex, non-stationary, and imbalanced nature of network traffic data. This framework addresses these challenges by operating in two distinct stages. The first stage incrementally models normal traffic patterns using lightweight regression and time-aware features, continuously adapting to evolving network dynamics. The second stage then analyzes the residuals from these predictions, combining them with contextual indicators to accurately identify genuine service-affecting failures. The framework is designed for fully online operation with low computational overhead, allowing for continuous adaptation. Evaluations show that this two-stage architecture outperforms other models across various metrics, achieving the best precision-recall trade-off, highest recall, F1-score, and AUC while maintaining acceptable false positive rates. This highlights the critical role of explicit residual decomposition in achieving reliable failure detection in real-time mobile network data streams.