Predicting Mobile Users Traffic and Access-Time Behavior Using Recurrent Neural Networks

Predicting mobile users' web-access behavior can have substantial impacts on resource allocation and cost reduction for wireless networks. Therefore, we propose a machine learning platform to forecast the web traffic and access time of mobile users. Based on the observation, the traffic patterns exhibit complex dependency on time, location, and popularity of webpages. Thus, a Recurrent Neural Network (RNN) with Long Short Term Memory (LSTM) is developed based on distinct engineered features to learn and predict users' web browsing activities. Then, to forecast the future access-time of mobile users, we propose a Self-exciting Memory Neural Network (SMNN). The access activities are modeled as self-exciting point processes, and intensities are adopted for prediction. Moreover, we extend the proposed predicting framework to cell towers. To cope with the diversity of traffic at the cell tower, we resort to clustering methods to group the similar users of each tower. Then, we develop an LSTM model for each cluster separately to predict the web domain traffic activities for the cell tower. Finally, we show that our proposed models outperform the baseline prediction models based on cellular networks dataset. We also show that for the cell tower access prediction task, the clustering method can significantly improve the prediction accuracy.