Current neural network ensemble methods usually generate accurate and diverse component networks by disturbing the training data, and therefore achieve strong generalization ability. In this paper, causal discovery is employed to discover the ancestor attributes of the class attribute on the results of the sampling process. Then, component neural networks are trained on the samples with only the ancestor attributes being used as inputs. Thus, the mechanism of disturbing the training data and the input attribute is combined to help generate accurate and diverse component networks. Experiments show that the generalization ability of the proposed method is better than or comparable to that of the ensembles generated by some prevailing methods.