Abstract
Coupled tensor decomposition reveals the joint structure of multi-modal data by incorporating a priori knowledge from latent coupled factors. With the permutation-invariance property, tensor ring (TR) decomposition ensures the uniformity of decomposed factors and mode attributes, which can result in powerful representation capability in some multidimensional data processing applications. In this paper, we propose the low rank coupled TR model for completing missing entries of tensors with shared latent factors. The proposed model enables flexible coupling among an arbitrary number of tensors and can be efficiently solved via an alternating least squares algorithm. The excess risk bound for this model shows the theoretical performance enhancement in comparison with other coupled nuclear norm based methods. The analysis is validated on synthetic data, and experimental results on several real datasets demonstrate its superiority over the state-of-the-art methods.
Funding Information
  • National Natural Science Foundation of China (W2412085)
  • Innovation Research Foundation of National University of Defense Technology (202501-YJRC-XX-021)

This publication has 45 references indexed in Scilit: