[Paper Review] Low-Rank Hankel Tensor Completion for Traffic Speed Estimation

Xudong Wang, Yuankai Wu+2 more|arXiv (Cornell University)|May 21, 2021

Tensor decomposition and applications26 references4 citations

TL;DR

This paper proposes a data-driven, model-free approach for traffic speed estimation using low-rank Hankel tensor completion. By applying spatiotemporal delay embedding to transform sparse trajectory data into a fourth-order Hankel tensor, the method leverages low-rank structure and truncated nuclear norm minimization via ADMM to accurately recover global and local spatiotemporal patterns, outperforming existing methods even under 88% data sparsity.

ABSTRACT

This paper studies the traffic state estimation (TSE) problem using sparse observations from mobile sensors. Most existing TSE methods either rely on well-defined physical traffic flow models or require large amounts of simulation data as input to train machine learning models. Different from previous studies, we propose a purely data-driven and model-free solution in this paper. We consider the TSE as a spatiotemporal matrix completion/interpolation problem, and apply spatiotemporal delay embedding to transform the original incomplete matrix into a fourth-order Hankel structured tensor. By imposing a low-rank assumption on this tensor structure, we can approximate and characterize both global and local spatiotemporal patterns in a data-driven manner. We use the truncated nuclear norm of a balanced spatiotemporal unfolding -- in which each column represents the vectorization of a small patch in the original matrix -- to approximate the tensor rank. An efficient solution algorithm based on the Alternating Direction Method of Multipliers (ADMM) is developed for model learning. The proposed framework only involves two hyperparameters, spatial and temporal window lengths, which are easy to set given the degree of data sparsity. We conduct numerical experiments on real-world high-resolution trajectory data, and our results demonstrate the effectiveness and superiority of the proposed model in some challenging scenarios.

Motivation & Objective

• To address the challenge of traffic state estimation (TSE) with sparse, incomplete data from mobile sensors, especially when physical models or large simulation datasets are unavailable.
• To develop a purely data-driven, model-free framework that captures both global and local spatiotemporal correlations in traffic speed data.
• To reduce reliance on prior physical assumptions or large-scale training data by exploiting inherent low-rank structure in spatiotemporal data.
• To improve imputation accuracy in extreme missing data scenarios, such as those with up to 88% missing values, by enhancing local and global pattern recovery.

Proposed method

• Transforms the incomplete spatiotemporal traffic speed matrix into a fourth-order Hankel tensor using two-way spatiotemporal delay embedding (Hankelization).
• Applies a balanced spatiotemporal unfolding of the tensor to enable efficient low-rank approximation via the truncated nuclear norm.
• Uses the Alternating Direction Method of Multipliers (ADMM) to solve the tensor completion problem with an efficient, iterative optimization algorithm.
• Imposes a low-rank constraint on the unfolded tensor to capture both global trends and local consistency in traffic dynamics.
• Employs only two hyperparameters—spatial window length τs and temporal window length τt—based on data sparsity and computational constraints.
• Utilizes truncated nuclear norm minimization to approximate tensor rank, enabling robust and efficient completion without full SVD computation.

Experimental results

Research questions

• RQ1 Can a purely data-driven, model-free approach effectively estimate traffic speed from sparse mobile sensor data without relying on physical traffic flow models?
• RQ2 How well can low-rank Hankel tensor completion recover complex spatiotemporal patterns such as shockwaves and congestion transitions in highly incomplete data?
• RQ3 What is the impact of spatiotemporal delay embedding on preserving local and global correlations in traffic speed data?
• RQ4 How does the proposed method compare to state-of-the-art TSE methods in terms of accuracy and computational efficiency under extreme data sparsity?
• RQ5 Can the truncated nuclear norm on balanced unfolding provide a computationally efficient and effective alternative to full tensor rank approximation?

Key findings

• The proposed STH-LRTC method achieves superior performance over MFTV and ASM in recovering shockwaves and small-scale traffic transitions, especially in high-missingness scenarios (e.g., 88% missing values).
• STH-LRTC successfully recovers small shockwaves that are blurred or missed by ASM, demonstrating enhanced local pattern preservation.
• The method reduces running time to 96.3 seconds (±5.4) on NGSIM data with 5% trajectories, significantly outperforming STH-SNN (290.5s) and STH-Tucker (288.7s), due to efficient single SVD per iteration.
• The low-rank structure effectively captures global trends, while delay embedding enhances local consistency, enabling accurate recovery of complex traffic dynamics.
• The two hyperparameters (τs and τt) are sufficient to control the trade-off between model sensitivity and computational cost, making the method practical for real-world deployment.