The field of software engineering data mining has become increasingly interested in cross-project defect prediction (CPDP), which uses defect codes from other projects to develop prediction models, addressing the problem of insufficient data in the model construction process. However, the source and target projects usually have large differences in data distribution, which reduces the prediction performance. Based on the learning idea of instance migration, the distribution of target project features is changed to be close to the distribution of source project features, thus improving the performance of cross-project defect prediction. Specifically, the cross-project software defect prediction is investigated by combining software metric features with instance migration methods. First, we analyze the difficulty of class imbalance in the training set and extract features from the instances in the training set. Then the idea of improved AdaBoost is proposed. Finally, the algorithm based on instance migration is proposed and experimented, and the result analysis shows the effectiveness and feasibility of the method.