A Weighted K-hyperline Clustering Algorithm for Mixing Matrix Estimation on Bernoulli-Gaussian Sources

In this paper, we presents an augmented K-means clustering-weighted hyper line clustering(K-WHLC) approach to solve the problem of mixing matrix estimation for Bernoulli-Gaussian sources. This algorithm employs the K-means clustering method as the stage of initialization and then uses a recursive weighted approach to robustly localize the direction of hyper lines with the PCA technology. Furthermore, a valid probabilistic criteria is proposed to detect true vectors of basis matrix from the hyper lines set. The advantage of weighting strategy lies in that it can suppress the effect of outliers and strengthen the precision of algorithm. A series of numerical simulations demonstrate its high performance on the task of mixing matrix estimation under the medium and large-scale cases.