Call Detail Record Forecasting: A Multitask Learning Architecture Using Deep Learning Networks for Mobile Traffic Forecasting
This repository provides the code base used to publish a research paper in IEEE PIMRC 2017[1], which is an early work on deep learning-based mobile traffic forecasting.
[1] C.-W. Huang, C.-T. Chiang, and Q. Li, “A Study of Deep Learning Networks on Mobile Traffic Forecasting,” in IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), Montreal, Canada, Oct. 2017.
Abstract—With evolution toward the fifth generation (5G) cellular technologies, forecasting and understanding of mobile Internet traffic based on big data is the foundation to enable intelligent management features. To take full advantage of machine learning, a more comprehensive investigation on a mobile traffic dataset with the latest deep learning models is desired. Therefore, a multitask learning architecture using deep learning networks for mobile traffic forecasting is presented in this work. State-of-the-art deep learning models are studied, including 1) recurrent neural network (RNN), 2) three-dimensional convolutional neural network (3D CNN), and 3) combination of CNN and RNN (CNN-RNN). The experiments reveal that CNN and RNN can extract geographical and temporal traffic features respectively. Comparing with either deep or non-deep learning approaches, CNN-RNN is a reliable model leading in all tasks with 70 to 80% forecasting accuracy.