Assessing spatiotemporal correlations from data for short-term traffic prediction using multi-task learning

Traffic flow prediction is a fundamental problem for efficient transportation control and management. However, most current data-driven traffic prediction work found in the literature have focused on predicting traffic from an individual task perspective, and have not fully leveraged the implicit knowledge present in a road-network through space and time correlations. Such correlations are now far easier to isolate due to the recent profusion of traffic data sources and more specifically their wide geographic spread. In this paper, we take a multi-task learning (MTL) approach whose fundamental aim is to improve the generalization performance by leveraging the domain-specific information contained in related tasks that are jointly learned. In addition, another common factor found in the literature is that a historical dataset is used for the calibration and the assessment of the proposed approach, without dealing in any explicit or implicit way with the frequent challenges found in real-time prediction. In contrast, we adopt a different approach which faces this problem from a point of view of streams of data, and thus the learning procedure is undertaken online, giving greater importance to the most recent data, making data-driven decisions online, and undoing decisions which are no longer optimal. In the experiments presented we achieve a more compact and consistent knowledge in the form of rules automatically extracted from data, while maintaining or even improving, in some cases, the performance over single-task learning (STL).Peer Reviewe