The invention discloses a software defect prediction method. According to the method, a sample with a class label and a sample without a class label are processed together, semi-supervised learning is utilized in the laplacian eigenmaps (LE), the LE method is improved, and meanwhile, in order to avoid that samples different in classification are mapped to a small low-dimensional neighbourhood, especially that defect samples are mapped to a defect-free sample neighbourhood, cost-sensitive information is introduced when the sample point distance is calculated through an LE algorithm. On the basis, the LE mapping precision is improved, and the discriminating performance of discriminating performance can be effectively improved through the method. The invention further provides a software defect prediction system. When the software defect prediction method and system are applied to an NASA database, it is proved through experiments that the effectiveness of the method is improved, and compared with other comparative methods, the classification performance of the method is improved to a certain extent.