Evaluating Recurrent Neural Networks for prediction of Multi- Variate time series VoIP metrics

This work examines the potential of predictive analysis to characterize the behavior of VoIP traffic in a mobile environment where approximately 40, 000 packets of voice traffic have been collected, processed, and analyzed. Starting with the construction of 6 specific QoS/QoE metrics extracted from a VoIP measurement campaign in an LTE-A environment, we face the problem of predicting the behavior of such metrics across time. A preliminary stage involves estimating a Vector AutoRegressive (VAR) model to capture correlations among the involved time series. This stage also involves statistical checks, such as stationarity and residual autocorrelations, in order to build a consistent model to be used for prediction. In the second stage, we employ a set of recurrent neural networks (simple RNN, LSTM, and GRU) to predict the behavior of selected QoS/QoE metrics. This choice is motivated by the fact that such techniques are able to handle temporal sequences, owing to their cell memory structure. Then, the employed techniques are contrasted in terms of both their offered performance and required computational time. Results provide valuable insights for constructing realistic traffic models (not artificially simulated ones) and useful information for network providers looking to optimize their resources based on usage patterns.