Linear and Non-Linear Spatio-Temporal Input Selection In Wireless Traffic Networks Prediction using Recurrent Neural Networks

Abstract

For the optimization of computer networks with high bandwidth requirements, it is necessary to predict the traffic of the wireless network. Its goal is to reduce maintenance costs and improve internet services. Feature selection is a major issue in multivariate time series (MTS) spatio-temporal modeling. Another problem is the dependency between input features, time lags, and spatial factors, so an appropriate model is needed. This study aims to provide solutions to two problems. The first is to improve a feature extraction and selection process in spatio-temporal MTS data for relevant features using Detrended Partial Cross-Correlation Analysis (DPPCA) and nonredundant features associated with linear using Pearson's correlation (PC) filters and non-linear associations using Symmetrical Uncertainty (SU) and a combination of both PCSUF. The second is to develop a spatiotemporal framework model using recurrent neural networks (RNNs) to get better performance than the traditional model. These methods are combined and tested using a data set of cellular networks with one hour intervals during November in three locations. Testing the effectiveness of the feature selection technique showed that 27.6% of the total extracted features were. The forecasting model with the DPCCA-SU-RNN combination method is the best performance by having RMSE = 380.7, R2 = 97% and MAPE = 10%.