Search papers, labs, and topics across Lattice.
The paper introduces TUBO, a machine learning framework designed for reliable network traffic demand matrix (DM) forecasting that addresses the limitations of existing deep learning models in handling traffic fluctuations. TUBO uses a burst processor to predict traffic spikes and a model selector to dynamically choose the optimal forecasting model based on input characteristics and uncertainty. Experiments on real-world datasets (Abilene, GEANT, CERNET) demonstrate that TUBO improves forecasting accuracy by up to 10x and achieves 94% accuracy in burst occurrence prediction, leading to significant throughput gains in proactive traffic engineering.
Network traffic forecasting gets a reliability boost: TUBO's dynamic model selection and burst prediction deliver up to 10x accuracy gains and 9x throughput improvements in real-world networks.
Network operation optimization through traffic forecasting holds great promise but faces key challenges due to the complex and bursty nature of traffic patterns. While deep learning models outperform traditional statistical methods in time series forecasting, they remain unreliable for network traffic due to limited adaptability to traffic fluctuations and diverse patterns. We present TUBO, a novel machine learning framework tailored for reliable traffic demand matrix (DM) forecasting. TUBO comprises two core components: a burst processor that isolates and predicts sudden traffic spikes, and a model selector that dynamically chooses the most suitable forecasting model from a diverse pool based on input characteristics and uncertainty estimation. This enables TUBO to deliver accurate, robust, and uncertainty-aware predictions. Evaluations on three real-world datasets (Abilene, GEANT, and CERNET) show that TUBO improves forecasting accuracy by up to $10 \times$ and achieves $\mathbf{9 4 \%}$ accuracy in burst occurrence prediction. As a downstream application, we apply TUBO to proactive traffic engineering (TE) and demonstrate throughput gains of $\mathbf{9} \times$ and $\mathbf{3} \times$ over reactive TE and the best prior proactive TE approach, respectively.