The invention belongs to the technical field of software defect prediction, and particularly relates to a software defect number predicting method based on feature selection and ensemble learning. For solving the problems that in prediction of the software defect number, irrelevant module features damage the performance of a defect prediction model, regression models have different prediction abilities, an optimal regression algorithm cannot be selected and the like, the method comprises the following steps: filtering irrelevant and redundant features by using a feature selecting method based on a package mode; then using six different regression algorithms including a linear regression algorithm, a ridge regression algorithm, a decision tree regression algorithm, a gradient boosting regression algorithm, a nearest neighbor regression algorithm and a multilayered sensor regression algorithm, an ensemble learning technology is used, and a comprehensive regression model is constructed according to data instances of which features are screened. Compared with a single regression model, the software defect number predicting method based on feature selection and ensemble learning has the characteristic that the accuracy of software defect number prediction is improved.