Abstract:
|
Multifold data structures are generally stored in high-dimensional objects defined as nth-order tensors. Candecomp/Parafac model (CP) can be used for modelling 4th order tensors. The application of these techniques is, however, quite limited due to procedural complexity and interpretational issues. The elevated number of degrees of freedom of CP model, however, can raise questions on its wide applicability especially in case of large complex data-sets. Extremely slow convergence as well as different types of degenerate results can occur while estimating parameters. Efforts aimed to address these issues connected with computational efficiency and accuracy are still mainstream in the associated field of research. This work aims at addressing these difficulties through estimates with a double optimization procedure algorithm in which values are first estimated by Self-Weighted TriLinear Decomposition algorithm and then refined through standard Alternating Least Squares steps.
|