This paper adopts a Bayesian approach to learn an optimal nonlinear classifier that is relevant to the classification task of semisupervised problems. The approach uses a prior weight to emphasize on the importance of class, which acts as a parameter of the likelihood function for both labeled and unlabeled data. We derive an expectationmaximization (EM) algorithm to compute maximum likelihood point estimate. Experimental results demonstrate appropriate classification accuracy on both synthetic and benchmark data sets.