Context-aware road travel time estimation by coupled tensor decomposition based on trajectory data

Urban road travel time estimation and prediction on a citywide scale is a necessary and important task for recommending optimal travel paths. However, this problem has not yet been well addressed: most existing approaches face serious data sparsity issues, e.g., lack of sensor data on several road segments; and it is a difficult task to capture context patterns around the road and incorporate context-aware information into travel time estimation models. Because of this, we propose to utilize trajectory data to model road travel times as this type of data covers more urban road segments than data from traditional traffic monitoring systems. Moreover, the trajectory itself has involved both travel times and the context of road congestion. A general framework for context-aware road travel time estimation (CARTE) is then put forward. Specifically, we adopt a third-order tensor to model spatiotemporal road travel times by setting the congestion level as a third dimension. By incorporating another context-aware information, namely points of interest (POI), a coupled tensor decomposition algorithm is proposed to fill in missing data. Eventually, we propose an algorithm to calculate an ultimate two-dimensional (spatial and temporal) travel time matrix by weighting the congestion probabilities of each congestion level. The effectiveness of the CARTE was validated on two real-world datasets and it was compared to the state-of-the-art methods. The experimental results demonstrate that the proposed travel time prediction approach always achieves the best performance in terms of accuracy with different data sparsity and prediction horizons. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes a context-aware road travel time estimation framework using trajectory data and third-order tensor modeling, combined with congestion levels and points of interest information to fill missing data and calculate the final travel time matrix. The effectiveness of the method was validated on two real-world datasets, showing superior accuracy performance compared to state-of-the-art methods.