On Extracting the Spatial-Temporal Features of Network Traffic Patterns: A Tensor Based Deep Learning Model

Recently, the Artificial Intelligence (AI) technology is widely employed in both academia and industry. Few researches on employing deep learning for network traffic control and wireless resource management have emerged as a new research direction in the communication networking area. However, how to deploy the deep learning structure in a universal way to suit for common network applications and how to format the training data still remain as a formidable research challenge. Furthermore, whether and why the deep learning structure is efficient in contrast with that of the shallow learning model, from the perspective of networking applications, has not been investigated well in the literature. In this paper, we address these issues, and propose a matrix and tensor based spatial-temporal training data format. Our proposal can be regarded as a universal characterization of network traffic patterns. Furthermore, a deep Convolutional Neural Network (CNN) structure is constructed to fit the proposed training data format of the corresponding tensor space. The performance of our envisioned tensor based deep learning model is further analyzed by comparing with the shallow learning model. Computer based simulation results demonstrate that our proposal achieves significant improvement in terms of both training accuracy and network performance.