Low-Rank Autoregressive Tensor Completion

Low-rank autoregressive tensor completion for spatiotemporal traffic data imputation. (IEEE TITS'22)

@article{chen2022low,
  title={Low-rank autoregressive tensor completion for spatiotemporal traffic data imputation},
  author={Chen, Xinyu and Lei, Mengying and Saunier, Nicolas and Sun, Lijun},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={23},
  number={8},
  pages={12301--12310},
  year={2022},
  publisher={IEEE}
}

Features

• Highlights

  • Present a missing data imputation approach that utilizes the day dimension of traffic data.
  • Build a tensor completion task through truncated nuclear norm minimization.
  • Consider univariate autoregressive process along the temporal dimension to characterize the time series trends.

• New features in the repo

  • Use conjugate gradient to solve the linear equations, intead of least squares in the transdim repo.
  • Include both CPU (with import numpy as np) and GPU (with import cupy as np) implementation.

Datasets

• Seattle freeway traffic speed dataset: This dataset contains freeway traffic speed from 323 loop detectors on the freeway network with a 5-minute time resolution (i.e., 288 time steps per day) over the first four weeks of January, 2015 in Seattle, USA. The processed data is of size 323-by-8,064 in the form of multivariate time series matrix. Alternatively, introducing the day dimension would lead to a tensor data of size 323-by-288-by-28. [tensor.npz]
• Portland highway traffic volume dataset: This dataset is collected from the highway network of the Portland-Vancouver Metropolitan region, which contains traffic volume from 1,156 loop detectors with a 15-minute time resolution (i.e., 96 time steps per day) in January, 2021. The processed data is of size 1,156-by-2,976 in the form of multivariate time series matrix. Alternatively, introducing the day dimension would lead to a tensor data of size 1,156-by-96-by-31. [volume.npy]

Supported by