Modeling with Insufficient Data to Increase Prediction Stability

This article proposes a procedure for small sample regression, systematically using the concept of robust Bayesian inference and a contaminated prior. The approach explores the possible domain of population information and attempts to estimate regression parameters further. A data augmentation step included in the procedure works to enlarge the original small data set by adding new data to it. It follows that when the expectation-maximization (EM) algorithm is used to output the hypothesis, approximating the true (but unobservable) parameters based on the enlarged data set. Both the augmented data set and the maximum likelihood estimate used are generated, based on the implementation of contaminated priors. The experiments show that the proposed procedure can effectively lower the mean squared error when modeling.