A Deep Learning Framework with Spatial-Temporal Attention Mechanism for Cellular Traffic Prediction

Abstract

Currently, traffic prediction in cellular communication has become an important way to relieve traffic congestion, and then guarantee users' quality of experience (QoE) in multimedia service. However, when concerning traffic prediction in the context of big data, traditional deep learning models only have limited memory lengths and thus do not satisfy high prediction accuracy demand. In this paper, we design a new deep learningbased framework, which is built on our proposed novel spatialtemporal attention mechanism, to further increase the prediction accuracy of cellular traffic resource. Specifically, we firstly design a big data hierarchical architecture with four levels to extract meaningful information. Then, considering that cellular traffic resource from base stations characterizes spatial-temporal dependencies, we integrate conventional temporal attention mechanism for target area with that from spatial neighbor areas, proposing a novel spatial-temporal attention mechanism. Finally, we utilize this spatial-temporal attention mechanism based LSTM (STaLSTMs) to predict cellular traffic resource from base stations. Experimental results demonstrate that our proposed framework has better performances in cellular traffic prediction than other competing models. Importantly, in some application scenarios, this framework can also maintain high accuracy but only with 60% data.