A network traffic prediction model based on reinforced staged feature interaction and fusion

With the increasingly intelligent services provided by the Internet, users’ requirements are further improved in communication quality. Real-time and accurate prediction of network traffic plays an important role in network resource allocation, abnormal traffic detection and other works. However, existing single or combined prediction methods cannot model its complex non-linear and spatial–temporal dependence. To address this problem, we propose a novel prediction model titled “reinforced staged spatial–temporal feature interaction and fusion (RSTIF)”. Specifically, we model the dynamic spatial dependence of the traffic through diffusion convolution. And then we deploy a staged feature interaction and fusion module. Spatial and temporal feature extractors are designed to cooperate with the interaction module. In the interaction process of feature integration-feedback, the complementary of temporal and spatial features are fully utilized, and the spatial–temporal dependence is modeled. We conduct extensive experiments on three real-world traffic datasets. The experiment results demonstrate that our model achieves state-of-the-art performance and is superior to the existing prediction models.