Segmentation-driven incremental learning for accurate network traffic prediction

Abstract

This study proposes a novel data segmentation method for incremental learning in network traffic prediction, leveraging change points in network configuration data (e.g., the number of users and network equipment). Isolating high-variance segments improves the incremental learning performance. Existing methods such as the PELT algorithm in ruptures face challenges in isolating high-variance segments and have the disadvantage of high computational costs. In contrast, the proposed method efficiently identifies high-variance segments by leveraging network configuration data as segmentation criteria. This approach not only circumvents the computational costs associated with parameter tuning but also facilitates more effective isolation of high-variance segments, leading to improved segmentation accuracy. Experiments show an average MSE of 1.799, outperforming baseline methods (No Segmentation: 2.846, RPT: 2.653) and enhancing prediction accuracy.