Evaluating Drift Detection Methods for Failure Prediction in 5G Network Resource Management
Lucas M Matni Bezerra, Adriano Madureira dos Santos, Glauco Estácio Gonçalves, Marcos César da Rocha Seruffo · 2025
In dynamic cloud environments, ensuring accurate and adaptive resource allocation is essential to maintain 5G network performance. This study compares concept drift detection strategies to improve fault tolerance and resource efficiency using the 5G3E dataset, which encompasses abrupt, gradual, and recurring changes in CPU and network demand. An LSTM model was deployed to predict resource usage, and three drift detection techniques—Page-Hinkley (PHT), ADWIN, and KSWIN—were applied to the prediction residuals. Experimental results show that PHT captured 50% of the CPU drifts and 66% and KSWIN, conversely, detected every drift event (100% recall) at the expense of numerous false positives. These findings highlight the trade-off between sensitivity and precision in dynamic 5G scenarios, underscoring the need to select a drift detection method based on acceptable levels of missed detections versus false alarms.