Low-Rank Tucker Approximation of a Tensor From Streaming Data

摘要: This paper describes a new algorithm for computing low-Tucker-rank approximation of tensor. The method applies randomized linear map to the tensor obtain sketch that captures important directions within each mode, as well interactions among modes. can be extracted from streaming or distributed data with single pass over tensor, and it uses storage proportional degrees freedom in output Tucker approximation. does not require second although exploit another view compute superior provides rigorous theoretical guarantee on error. Extensive numerical experiments show produces useful results improve state art decomposition.