With the gradual penetration of machine learning technology into various fields, software testing in the software development process is very important. Software defect prediction faces class imbalance problem and accuracy issue. This paper proposes a supervised learning-based software prediction method for solving these two core problems. The method adopts sample balancing technique, combined with synthetic minority over-sampling technique(SMOTE) and edited nearest neighbor(ENN) algorithm, to test local weight learning(LWL), J48, C4.8, random forest, Bayes net(BN), multilayer feedforward neural network(MFNN), supported vector machine(SVM), and naive Bayes key(NB-K). These algorithms are applied to three different datasets (KK1, KK3 and PK2) in the NASA database and their effects are compared and analyzed in detail. The results show that the random forest model combining SMOTE and ENN exhibits high efficiency and avoiding overfitting in dealing with class imbalance problems, which provides an effective way to solve the problem in software defect prediction.