Automated learning of factor analysis with complete and incomplete data

In the application of the popular maximum likelihood method to factor analysis, the number of factors is commonly determined through a two-stage procedure, in which stage 1 performs parameter estimation for a set of candidate models and then stage 2 chooses the best according to certain model selection criterion. Usually, to obtain satisfactory performance, a large set of candidates is used and this procedure suffers heavy computational burden. To overcome this problem, a novel one-stage algorithm is proposed in which parameter estimation and model selection are integrated in a single algorithm. This is obtained by maximizing the criterion with respect to model parameters and the number of factors jointly, rather than separately. The proposed algorithm is then extended to accommodate incomplete data. Experiments on a number of complete/incomplete synthetic and real data reveal that the proposed algorithm is as effective as the existing two-stage procedure while being much more computationally efficient, particularly for incomplete data.