We introduce a mixture-of-experts technique that is a generalization of mixture modeling techniques previously suggested for semi-supervised learning. We apply the bias-variance decomposition to semi-supervised classification and use the decomposition to study the effects from adding unlabeled data when learning a mixture model. Our empirical results indicate that the biggest gain from adding unlabeled data comes from the reduction of the model variance, whereas the behavior of the bias error term heavily depends on the correctness of the underlying model assumptions.