Random Vcol [Albright2006] is inexpensive initialization method for nonnegative matrix factorization. Random Vcol forms an initialization of each column of the basis matrix (W) by averaging p random columns of target matrix (V). Similarly, Random Vcol forms an initialization of each row of the mixture matrix (H) by averaging p random rows of target matrix (V). It makes more sense to build the basis vectors from the given data than to form completely random basis vectors, as random initialization does. Sparse matrices are built from the original sparse data.
Method’s performance lies between random initialization and centroid initialization, which is built from the centroid decomposition.
Return initialized basis and mixture matrix. Initialized matrices are of the same type as passed target matrix.