The lack of traffic data due to factors such as sensor failure and communication failure seriously restricts the development and application of ITS. How to recover missing data accurately and efficiently has become a key issue for ITS. In recent years, the method of LRTC (Low Rank Tensor Completion) has been widely used in traffic data completion. This article will introduce several recent literature on traffic data completion. Criticism and correction are welcome!
1、LRSSRTC（2022）
文献：A Novel Spatiotemporal Data Low-Rank Imputation Approach for Traffic Sensor Network
Journal: IEEE Internet of Things Journal
Authors: Xiaobo Chen, Shurong Liang, Zhihao Zhang, Feng Zhao
Abstract: The Internet of Things (IoT) has great potential to transform the transportation industry by improving passenger experience, safety, and efficiency. However, the spatio-temporal data collected by traffic sensor networks are often affected by missing values (MV), which affect the overall performance of the system. Therefore, the accurate recovery of MVs is crucial for the successful application of IoT in transportation. In this paper, we propose a new imputation model for MVs by integrating low-rank tensor completion (LRTC) and sparse self-representation into a unified framework. In this way, the global multidimensional correlation and sample self-similarity can be well utilized for imputation. To solve the proposed model, a complex solution algorithm is proposed in this paper following the principle of alternating direction methods of multipliers (ADMMs). Importantly, each step in ADMM can be effectively implemented by analyzing the problem structure. Furthermore, in order to select appropriate parameters for the model, this paper proposes an improved harmony search heuristic algorithm based on the dual harmonic generation strategy, which fully considers the information contained in the current harmony memory. Experiments are conducted on two real-world traffic data to evaluate the proposed method. The results show that the method significantly improves the imputation performance compared to classical matrix/tensor completion algorithms and other competing algorithms.
Main idea: use weighted SNN to model global low-rank structure, and use sparse self-representation to model local characteristics (periodicity).
2、TBTC（2023）
文献：Transforms-based Bayesian Tensor Completion Method for Network Traffic Measurement Data Recovery
期刊：IEEE Transactions on Network Science and Engineering
Authors: Zecan Yang, Laurence T. Yang, Lingzhi Yi, Xianjun Deng, Chenlu Zhu, Yiheng Ruan
Abstract: Network traffic measurement is regarded as the cornerstone of next-generation network systems. Its purpose is to monitor network traffic and provide data support for traffic engineering. Therefore, it is particularly important to monitor traffic data from the perspective of the entire network. However, the surge of network services has led to explosive growth of network traffic, which poses a major challenge to the measurement of network traffic. Therefore, how to infer the entire network traffic from partial traffic data is extremely important. This paper proposes a transform-based Bayesian tensor completion (TBTC) method to infer network traffic data. First, the heterogeneous network traffic data with missing entries is organized into observation tensors according to the time dimension and other attributes; second, the side-slice sparsity of factor tensors is induced by using the sparse hierarchy prior, so that the fallopian rank of observation tensors can be estimated; Further, a variational Bayesian inference method for model learning is proposed, and an efficient update method is proposed. Finally, two computational examples of the tensor completion model based on linear transformations are implemented in experiments. Experimental results on two real network traffic datasets verify that the proposed method can recover network traffic data efficiently and accurately.
The main idea: Based on Bayesian theory, the missing tensor is expressed as the cosine transformation product of two factor tensors/tensor-tensor product of any reversible linear transformation, and then optimized and solved by TNN.
Cosine transform product:
tensor-tensor product of any invertible linear transformation
3、ManiRTD （2023）
文献：Manifold Regularized Tucker Decomposition Approach for Spatiotemporal Traffic Data Imputation
Source: arXiv:2305.06563v2 [stat.ML] 16 May 2023
Authors: Wenwu Gong, Zhejun Huang, Lili Yang
Abstract: Spatio-temporal traffic data interpolation (STDI), which estimates missing data from partially observed traffic data, is an unavoidable challenge in data-driven intelligent transportation systems (ITS). Due to the multidimensional and spatiotemporal properties of traffic data, we treat missing data imputation as a tensor completion problem. In the past decade, many studies have been based on tensor decomposition of STDI. However, how to exploit spatio-temporal correlation and core tensor sparsity to improve imputation performance remains to be solved. In this paper, we reshape rank-3/4-rank Hankel tensors and propose an innovative manifold-regularized Tucker Decomposition (ManiRTD) STDI model. Specifically, we represent perceptual traffic state data as 3/4th tensors by introducing a multi-way delayed embedding transformation. Then, ManiRTD improves the sparsity of the Tucker kernel with a sparse regularization term, and adopts the manifold regularization of the factor matrix and the temporal constraint term to characterize the spatio-temporal correlation. Finally, the ManiRTD model is characterized by a block coordinate descent framework under the alternating proximal gradient update rule. Numerical experiments are performed on the real-world spatiotemporal traffic dataset (STD). Our results show that the proposed model outperforms other factorization methods and more accurately reconstructs the STD under various missing scenarios.
The main idea: Use multi-way delayed embedding transformation to reshape the matrix to form a fourth-order Hankel tensor; use Tucker decomposition to mine long-term global trends, and capture short-term local trends through manifold regularization and Toeplitz regularization.
4、LETC（2023）
文献：Correlating sparse sensing for large-scale traffic speed estimation: A Laplacian-enhanced low-rank tensor kriging approach
期刊：Transportation Research Part C: Emerging Technologies
Authors: Tong Nie, Guoyang Qin, Yunpeng Wang, Jian Sun
Abstract: Traffic speed is the core of characterizing road network mobility. Many transportation applications rely on it, such as real-time navigation, dynamic route planning, congestion management, etc. Rapid advances in sensing and communication technologies have made traffic speed detection easier than ever. However, due to the sparse deployment of static sensors or the low penetration rate of mobile sensors, the detected velocities are incomplete and far from network-wide usage. Additionally, sensors are prone to erroneous or missing data for various reasons, and these sensors can get very noisy at speed. These shortcomings require efficient techniques to recover plausible estimates from incomplete data. In this work, we first identify the problem as a spatiotemporal kriging problem and propose a Laplacian-enhanced low-rank tensor completion (LETC) framework that is both low-rank and multidimensionally correlated properties for large-scale traffic velocity kriging under limited observations. Specifically, three types of velocity dependencies including temporal continuity, temporal periodicity, and spatial proximity are carefully selected and simultaneously modeled with three different forms of graph Laplacian, namely the temporal graph Fourier transform , Generalized Temporal Consistency Regularization and Diffusion Map Regularization. Then, we design an efficient solution algorithm to extend the proposed model to full network kriging through several efficient numerical techniques. By conducting experiments on two publicly available million-class traffic speed datasets, we finally draw conclusions and find that our proposed LETC achieves state-of-the-art kriging performance even at low observation rates, while being comparable to baseline methods. than saving more than half of the computing time. It also provides some insights into spatio-temporal traffic data modeling and kriging at the network level.
Main idea: On the basis of TNN, TGFT (describe time periodicity), DGR (describe spatial proximity), GTCR (describe time continuity) regularization items are introduced.
