Comparison of Network Traffic Forecasting Schemes

The network traffic prediction or forecasting problem has attracted tremendous interest from both academia and industry because traffic prediction is widely used as the fundamental data in establishing or in updating network infrastructure. Recently, numerous clever time-series analysis schemes based on machine learning techniques such as RNNs (Recursive Neural Networks) and self-attention architecture have been developed. However, most of the previous studies in the area of artificial intelligence (AI) do not deal with network traffic in their experiments. In this paper, we deal with the network traffic prediction problem using the domain knowledge accumulated in the field of networking. In particular, we define channels for network traffic and interpret temporal and spatial information from the real-world traffic dataset. Our experiments on the Abilene dataset show that the recent patch mixing model performs best in terms of prediction error. In addition, we observe that both RNN models and simple transformer-based models show inadequate performances.