Neural network architecture based on gradient boosting for IoT traffic prediction

Abstract Network traffic forecasting is an operational and management function that is critical for any data network. It is even more important for IoT networks given the number of connected elements and the real-time nature of many connections. This work presents a novel deep learning architecture applicable to this supervised regression problem. It is based on an additive network model formed by ‘learning blocks’ that are stacked iteratively following, in part, the principles of gradient boosting models. The resulting architecture is trained end-to-end using stochastic gradient descent. This new architecture has connections with residual, stacked and boosted networks, being different from any of them. Like residual networks, it shows excellent convergence behavior during training and allows for deeper models. It has a regularization effect similar to stacked models and presents excellent prediction results as gradient boosting models do. The building elements of the architecture are neural network blocks or learning blocks, that can be constituted by a sequence of simple fully connected layers or by more elaborate dispositions of recurrent and convolutional layers. The resulting architecture is a generic additive network (gaNet) applicable to any supervised regression problem. To obtain experimental results on a hard prediction problem, the model is applied to the forecasting of network traffic using IoT traffic volume real data from a mobile operator. The paper presents a comprehensive comparison of results between the proposed new model and many alternative algorithms, showing important improvements in terms of prediction performance metrics and training/prediction processing times.