International Conference on Emerging Technologies for Communications
Tensor Decomposition Based Method for Multiset Blind Source Separation
Lu Wang, Tomoaki Ohtsuki,
In this paper, we propose to apply the tensor theory to develop a novel structure for the blind source separation (BSS) technique that allows us to exploit the second-order statistic in terms of the tensor format. In our model, the tensor computing is regularized by a low-rank prior to exploring a stack of cross-covariance matrices along with temporal information. The approach in terms of limited storage can effectively encapsulate and condense the large-scale data into a compact format. In addition, it is beneficial to decompose the tensor format into multiple low-rank tensors of matrices, providing an efficient way for analyzing complex data. The experiment designed on synthetic and real-world biological data, and the proposed algorithm consistently presents more accurate normalized mean squared error (NMSE) than three other classical algorithms. The possible reasons are due to the separation criterion, which exploits not only the statistical independence within multiple datasets, but also dependence among the different data sets.