Imputation Algorithm

Another option is to complete the data matrix by iteratively imputing the missing values (see, e.g., [8]). Initially, the missing values can be replaced by zeroes. With completed data, PCA can be solved by eigenvalue decomposition of the covariance matrix. Now, the missing values are replaced using the product $\mathbf{A}\mathbf{S}$, PCA is applied again, and this process is iterated until convergence. This approach requires the use of the complete data matrix, and therefore it is computationally very expensive if a large part of the data matrix is missing.