Aiming at dealing with the problems of the data deficiency, data high dimensionality in software defect prediction (SDP), this paper proposes a novel three-stage defect prediction model. First we introduced the information flow algorithm (IFA) to do the causality analysis to choose the most representative feature subset. Then we utilized the nearest neighbor method in similarity measuring between the local dataset and the cross-project data, to construct a homogenous training dataset. Last we proposed a new Bayes classifier, which evolves the Naive Bayes classifier with the vibration of string based diffusion function (VSDF).I