Low-rank Tensor Autoregressive Predictor for Third-Order Time-Series Forecasting

Abstract: Recently, tensor time-series forecasting has gained increasing attention, whose core requirement is how to perform dimensionality reduction. Among all multidimensional data, third-order tensor is the most prevalent structure in real-world scenarios, such as RGB images and network traffic data. Previous studies in this field are mainly based on tensor Tucker decomposition and such methods have limitations in terms of computational cost, with iteration complexity of approximately $O(2n^3r)$, where $n$ and $r$ are the dimension and rank of original tensor data. Moreover, many real-world data does not exhibit the low-rank property under Tucker decomposition, which may fail the dimensionality reduction. In this paper, we pioneer the application of tensor singular value decomposition (t-SVD) to third-order time-series, which builds an efficient forecasting algorithm, called Low-rank Tensor Autoregressive Predictor (LOTAP). We observe that tensor tubal rank in t-SVD is always less than Tucker rank, which leads to great benefit in computational complexity. By combining it with the autoregressive (AR) model, the forecasting problem is formulated as a least squares optimization. We divide such an optimization problem by fast Fourier transformation into four decoupled subproblems, whose variables include regressive coefficient, f-diagonal tensor, left and right orthogonal tensors. The alternating minimization algorithm is proposed with iteration complexity of about $O(n^3 + n^2r^2)$, in which each subproblem has a closed-form solution. Numerical experiments show that, compared to Tucker-decomposition-based algorithms, LOTAP achieves a speed improvement ranging from 2 to 6 times while maintaining accurate forecasting performance in all four baseline tasks. In addition, LOTAP is applicable to a wider range of tensor forecasting tasks due to its more effective dimensionality reduction ability.