Deep Convolutional LSTM Network-based Traffic Matrix Prediction with Partial Information

Accurate prediction of the future network traffic plays an important role in various network problems (e.g. traffic engineering, capacity planning, quality of service provisioning, etc.). However, the modern network communication is extremely complicated and dynamic, which makes the tasks of modeling and predicting the network behavior very difficult. To this end, a common approach is to apply the traditional time series prediction techniques such as Autoregressive Integrated Moving Average or Linear Regression. Besides that, there are some studies exploiting Deep Learning techniques such as Restricted Boltzmann Machine or Recurrent Neural Network (RNN) to estimate the traffic volume. Although the prediction accuracy largely depends on the amount of historical data, measuring all the network traffic is impossible or impractical due to the monitoring resources constraints as well as the dynamics of temporal/spatial fluctuations of the traffic. Thus, the state-of-the-art proposals reveal poor performance regarding the traffic inference when lacking ground-truth input.In this paper, we propose a highly accurate traffic prediction algorithm by leveraging the Convolutional LSTM network (ConvLSTM), which is the integrated model of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network, for spatiotemporal modeling and estimating the future network traffic. We also propose a technique which exploits the RNN to correct the imprecise data in the input. To evaluate the proposed algorithm, we conduct extensive experiments using the Abilene dataset which contains the real network traffic trace. The experiment results show that our proposed approach outperforms the existing algorithms in terms of several metrics including error ratio, root mean square error, and coefficient of determination, in both one-step-ahead and multi-step-ahead prediction with partial information.