Adaptive Online Learning for Network Traffic Prediction

Accurate real-time prediction of network traffic is crucial for adaptive resource management and robust operation in modern telecommunications networks. However, rapidly varying patterns and missing data present significant challenges for standard forecasting approaches. Offline learning, which depends on static datasets and periodic retraining, often struggles to adapt quickly to such dynamic conditions. In contrast, Online Learning (OL), which updates models incrementally as new data arrives, can adjust more quickly to changes in network traffic patterns. In this study, we propose OL for network traffic prediction and compare it with offline learning for time series forecasting. Although OL adapts quickly to changing traffic, fixed learning rates may still slow convergence under shifting data. To address this, we apply Adaptive Learning Rate (ALR) methods that adjust step sizes automatically, improving stability and responsiveness. We evaluate two ALR approaches, Hypergradient Descent and MetaGrad, within the OL framework against the fixed-rate Adam optimizer, using SIX and CESNET for telecom traffic and Jena Climate for environmental data. Our experimental results show that OL consistently outperforms offline learning in responsiveness, while ALR methods further improve overall adaptability, enabling forecasting approaches that can operate effectively in real-time settings required by next-generation sixthgeneration (6 G) networks. Index Terms-Online machine learning, traffic prediction, time series forecasting, adaptive learning rate, MLP.