Traffic Matrix Prediction with Attention-based Recurrent Neural Network

Abstract

Traffic matrix (TM) shows the traffic volume of a network. Therefore, TM prediction is of great significance for network management. Attention mechanism has been successful in many sub-domains of machine learning, such as computer vision and natural language processing, and it performs particularly well on time series data. In this work, we first introduce attention mechanisms into the traffic matrix prediction field by proposing an attention-based deep learning model for traffic matrix prediction. This model is composed of two parts, encoder and decoder. We use a recurrent neural network (RNN) architecture as the encoder and our decoder has an attention layer and a linear layer. Attention mechanism allows the model to have better memory ability, so the model can concentrate on those important data regardless of distance. We also reduce the time consumption of our model using GPU-based parallel acceleration. Finally, we evaluate the effectiveness of our model on a real world TM dataset, and the results show our implementations on the proposed model perform better than the baseline models.