Traffic flow forecasting based on multitask ensemble learning

A new method for traffic flow forecasting based on multitask ensemble learning, which combines the advantages of multitask learning and ensemble learning, is proposed. Traditional traffic flow forecasting methods are a single task learning mode, which may neglect potential rich information embedded in some related tasks. In contrast to this, multitask learning can integrate information from related tasks for effective induction. Recent developments also witness the potential of ensemble learning for traffic flow forecasting. This paper devises a new method named MTLBag, a combination of multitask learning and a famous ensemble learning method bagging, for traffic flow forecasting. Using a neural network predictor, this paper first empirically shows the superiority of multitask learning over single task learning for traffic flow forecasting. Experimental results also indicate that the performance of MTLBag is statistically significantly better than that of the multitask neural network predictor, and that MTLBag outperforms a state-of-the-art method Bayesian networks.