Network Traffic Prediction Based on Spiking-Inspired Transformer

With the development of network functional scenarios and architecture, it poses new challenges to the performance improvement of the network, especially for the network traffic prediction. In this paper, a Spiking-inspired transformer, which adds spiking properties to enhance the attention mechanism of transformer, is proposed to improve the accuracy of network traffic prediction. Primarily, the original characteristics of traffic data is extracted by the multi-attention module layer, and the extracted data is encoded by the spiking-attention layer. By introducing LIF neurons, the spiking-attention layer is able to simulate the pulse-firing behavior of biological neurons to capture pulse properties and long-range dependencies in network time series data. Finally, the predict result is produced by the decoded and output layer. Through extensive experiments on network traffic datasets, such as GEANT network data, we demonstrate the superior performance of our proposed model compared to existing approaches, which shows advancement in traffics prediction.