Convergence acceleration of alternating least squares with a matrix polynomial predictive model for PARAFAC decomposition of a tensor. Shi, M., Zhang, J., Hu, B., Wang, B., & Lu, Q. In 2017 25th European Signal Processing Conference (EUSIPCO), pages 1115-1119, Aug, 2017.
Convergence acceleration of alternating least squares with a matrix polynomial predictive model for PARAFAC decomposition of a tensor [pdf]Paper  doi  abstract   bibtex   
In this paper, a matrix polynomial whose coefficients are matrices is first defined. Its predictive model, called as the Matrix Polynomial Predictive Model (MPPM), is then derived. When the loading matrices of a decomposed tensor in the Alternating Least Squares (ALS) are replaced by the predicted ones of the MPPM, a new ALS algorithm with the MPPM (ALS-MPPM) is proposed. Analyses show that the convergent rate of the proposed ALS-MPPM is closely related to the degree of the matrix polynomial. Namely, when an accelerative convergence rate is expected, the polynomial with a high degree is preferred. Although a high degree means a high possibility of prediction failure, a simple solution can be used to handle such failure. Moreover, the relationship between our ALS-MPPM and the existing ALS-based algorithms is also analyzed. The results of numerical simulations show that the proposed ALS-MPPM outperforms the reported ALS-based algorithms in the literature while the analytical results are verified.

Downloads: 0