MSAE: A Multitask Learning Approach for Traffic Flow Prediction Using Deep Neural Network

Abstract

Traffic flow prediction has been regarded as one of the key problems in Intelligent Transportation Systems. Neural networks are widely leveraged in traffic prediction task, but with the limitation of (1) single task learning, which would ignore shared information among traffic network; and (2) initialization problem that would directly affect the training efficiency. In this work, we propose a deep neural network based multitask learning approach for traffic flow prediction, called MSAE, which incorporates stacked autoencoders (SAE) for the neural network initialization, and jointly and adaptively predicts network-scale traffic flow with shared information. The experiments on real traffic data indicate that the proposed model, jointly considering shared information among different traffic flow prediction tasks and neural network initialization, is capable and promising of dealing with complex traffic flow forecasting with satisfying accuracy and effectiveness.